July 30, 2025 • 58 mins
In this event, Shane McAllister will cover how Albert Podusenko got introduced to MongoDB and in 2016 decided to start a company using MongoDB as primary database behind their server, handling everything: From model states and real-time event generation to maintaining episodes and storing historical data.In this show we'll talk about:
- How to make a product out of research.
- Walk through it's journey that culminated with the creation of Lazy Dynamics.
- Show how they utilize MongoDB with demos and examples.
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Hello, welcome to MongoDB Podcast Live.
I'm Shane McAllister. It's great to have you.
And today we're diving into the fascinating intersection of
robotics, machine learning and real world applications with the
true innovator. And to elaborate on the amazing
title of turning chaos into opportunity, please join me in
(00:20):
welcoming Albert, the CEO of Lazy Dynamics, a company
revolutionizing how systems react and adapt in dynamic
environments. Albert, you're very welcome to
the Mongo to Be podcast. Hi, Shane.
Yeah, thanks a lot for this introduction.
Happy to be here. Not at all.
It's great to have you. Thank you for joining.
(00:40):
I always start and I know we dida prep call on this.
I always start with our guest totell us a little bit about their
career path to date and how theygot here and in our
conversation, Albert, you know, that was incredibly interesting.
So I know we don't have all the time in the world, but maybe
perhaps give us a potted historyof of your background to date
(01:02):
and and how you ended up being the CEO of Lazy Dynamics.
Sure, sure. Yeah.
I'll try to compress it as much as possible.
But yeah, so my background, so basically I hailed from Far East
of Russia, a place where far from where I am based right now,
which is in the Netherlands. I completed my bachelor's degree
in the western part of Russia, like at Saint Petersburg State
(01:25):
University, where I was studyingapplied mathematics and control.
So upon graduation of my bachelor's, I, I think I was a
bit likely to get a scholarship for a two years master program
in, in Japan, in Kyoto, in computer science, where I dove
into self driving cars and driving aids research.
(01:47):
But it was odd. Well, it, it's still getting it,
it it's still odd. Or like, let's say it's WAVY,
right? It's like always five years
away. Back in 2016, it was quite hot
was coming in two years and and yeah, so I would I embark on
that research where we were developing essentially an an
assistant for elderly drivers. So that could predict their
(02:09):
intention to emergency brake before they even hit the the
braking pedal. And the research, I must say,
like progressed well, we even secured a patent.
But when we tested it with real drivers, my friends, we hit
quite a roadblock. So they basically hated the, the
system. They, they really didn't like
(02:30):
it, to put it mildly. And well, and the reason was
that it generated too many falsepositive.
So it was trained on the so-called average driver, right?
And like, well, now it's retrospectively it seems
obvious, but well, back in the days, like it wasn't.
So the, the problem with averagedriver, it doesn't exist.
(02:52):
So, so and then it led me to explore like adaptive systems,
adaptive technologies, how couldwe kind of leverage that prior
knowledge from many drivers thatwe've well that we've researched
right to, but then to fine tune itself in real time.
So how can we kind of have a system that can fine tune itself
(03:15):
in real time to the actual person behind the wheel.
And I think this eventually led me or guided to my PhD working
on adaptive hearing aids. And it, it may seem that it's
completely orthogonal problem, right?
Like where is like the driving cars or like where is the
hearing aid patients, right. And in a way there are different
(03:39):
problems, but what connects them?
There is no average hearing aid patient.
There is no average driver. And for both tasks we actually
require to have a system, or nowas they call it, an agent that
can adapt into the ever changingworld.
The acoustic environment changes, right?
(03:59):
They're all changes. And then also there's kind of a
sense of the personalisation, like we want to personalize the
hearing aid algorithm to basically work for a particular
human driver. Sorry, for a hearing aid
patient. And yeah, that's basically what
I've, I've been working on during my PhD.
And yeah, after in four years we've basically developed the
(04:23):
general toolbox, a solution as we call it, an engine that
enables this type of adaptive learning.
And well, about a year ago we spun out from our laboratory and
I guess, well, everyone in my team, Co founding team took nice
roles and like I was assigned asa CEOI would say, well, so
(04:47):
that's, that's kind of how it worked out.
Yeah. I love that assigned as ACEO.
So let's talk a little bit aboutthat.
I suppose maybe the team itself,the diverse academic background
that the team has that you've assembled in Lazy Dynamics and I
suppose in in the pathway of talking about that team talk a
(05:09):
bit about the driving force of the the spin out from from the
university then as well. Albert Right, right.
So I, I think the, the, the coreteam consists of five PhDs, so
each bringing, bringing criticalexpertise like we have our
mathematics guy, our coding expert, an engineer, and of
(05:31):
course the research lead. And I guess, and there is me
like serving as a as a CEO. So precisely because I'm not
excellent at any of those other specialized roles.
So yeah, but I'm. Sure, you're very much
underestimating yourself there now.
Yeah, that's what they say. But, but I, I think, I think I'm
definitely good at bringing it all together.
So that's, that's basically, I guess a connection point for all
(05:54):
that. But that, that background is
rather diverse, I would say. So we actually all came from
different masters and, and the bachelor's.
So some people were like doing mathematics, some people were
studying neuroscience. And then after all, it sort of
started to converge into one solid idea on creating this sort
(06:17):
of engine or software, software toolbox that merges many
different ideas to create, well,one solid framework to help to
enable, as we call it like now, adaptive intelligence.
OK. OK.
So from a timeline point of view, that team which as you
mentioned a blend of maths and engineering and coders etcetera
(06:40):
and the scientific knowledge behind it all that that was
about a year or more ago then was it Albert?
Yeah, I think so. So we we were incorporated 1 1/2
year ago, but we truly started to work I think like 2024.
And the reason for that was simple.
So that when we established the company, they were some of our
(07:00):
Co founders who are still doing PhDs like actually like they
were very close to dance which was not an option to do a drop
out. So which normally founders do.
But in our case, it was like just three more, three more
months that of course extended to half a year.
But nevertheless they finished now.
So now we're all business basically.
OK, wow, wow. So look in your own case, a big
(07:24):
a huge background there and variety.
And as you said, there are similarities between, you know,
the self driving and the hearingaid.
But tell us a little bit about what lazy dynamics then is is
focusing on. Yeah, I I think the best way
actually to go about it is like to to well to go a bit backward,
(07:44):
it's actually like so to to to start with a bit of well
observation about this kind of founding story.
So, and the interesting observation is actually about
this current AI systems. So everyone is now talking about
agents, about entities essentially that perceive their
environment, take autonomous actions to achieve certain
(08:05):
goals. And the problem we identified at
Lazy Dynamics and we kept identifying throughout our PhDs
is that most authentic systems that that are based on for
instance, LLMS which are fantastic technology with great
implementation and content creation outreach campaigns.
(08:26):
However, when you when you try to build, let's let's say agents
for mission critical sectors, say autonomous systems, smart
cities, finance or hearing aids.OK.
The partners we are working on working with, they aren't eager
to bet lol. Based agents that can
hallucinate and struggle to adapt quickly enough.
(08:49):
And this is kind of exactly whatwe provide with our engine.
So identic systems based on probabilistic modelling.
I will not go into much details,not to this conversation, but
they basically we try to make agents that adhere to racial
reasoning and adapt rapidly to changing conditions.
(09:10):
And just to give you like a really dangerous example, I
think so you can imagine, let's say like an ambulance rushing to
a hospital. Yeah.
OK. And an agent which is inside
this emergency car, right? So it needs to find the fastest
route. So it needs to continuously like
(09:30):
process real time traffic, day data, weather conditions,
emergency AOV, equal locations to dynamically update this
optimal path. So and so unlike the traditional
navigation systems, our agent, it adapts its decisions based on
the changing probabilities. So say you are in the junction
(09:51):
and like now you can take left or right and the system tells
you, OK, if you go left, it willbe 6 minutes, if you go right,
it will be 5 minutes. However, I am very uncertain
about if you actually check turnright.
Yeah, this 5 minutes can actually turn to be 15 because
there are like some conditions there.
So the agent actually tries to proactively suggest alternative
(10:15):
by considering this certainty around certain decisions that
it's going to make. So yeah.
OK, OK. I think that's a good real world
example, but I suppose you're how are you, you know, what's
the heart of the system there that you're building?
You mentioned the toolbox earlier as well too, you know,
(10:38):
how does it and I understand that example of an ambulance and
there's a driver hopefully in the ambulance.
We haven't got to that stage yetwhere we don't have the driver.
But obviously you're looking towards, you know, drones and
driverless cars and other thingsthat are making these decisions
on their own, right? So you the toolbox you're saying
that laser dynamics have built. How is?
(10:59):
How is that addressing that bigger problem right?
So the toolbox is called Riggs and FUR and it's basically an
engine that allows to build agents.
So what we do is essentially we provide tooling for users to
(11:21):
create a genetic systems by using some Lego blocks as you
would like adaptive Lego blocks that we call probability
distributions or like let's say functions and allow to well
essentially describe the environment where the agent
lives, right? So be it.
(11:42):
If we are talking about the drone, we describe through some
high level abstraction functions, OK, how's the
dynamics of the drone changes like say if we if I'm here at
the point XYZ and I want to go to XYZ TT1, right?
So how, how this dynamics work? And The thing is that some
(12:07):
people they know these dynamics already, so they can just
provide those. But sometimes it's not really
known like there are like some robotic system for which you
might not have the full description of this of the
internal dynamics or of the model of the world.
So we basically create this additional building blocks which
(12:27):
also encapsulate the parameters of this unknown functions.
And the idea here is that once the drone goes out in the world
to solve the task and the task, we can mean like like a
surveillance or or we can think of going just from one from
point A to point B. The drone needs to figure out
(12:48):
OK, what type of the actions or like controls its needs to apply
in order to reach the goal. At first it it might have to
well to learn a bit. It's dynamic, so it needs to do
exploration. So essentially what we do is
like what our SDK allows to to create is well, describe the
model of the world for the agentand then run the inference.
(13:12):
So basically then you press the button once you've specified
your agentic system, your goal. So once you run our SDK, it
allows, well, it provides you with the solution, sort of a
solver how to reach that certaingoal in the environment you
describe. OK, fascinating.
And I know we're going to see a demo of some of this later on
(13:35):
towards the end of our conversation.
So for our viewers joining in now with Albert and me chatting
away about what's very complex, you are going to get an
illustration of how this works and what's it entails as well
too. And, and for anybody joining the
live stream, if you've got any questions as we go through, you
may not have missed seen the beginning where I was saying and
(13:56):
some people have shouted out said who they are.
Do let us know where you're joining in from as well too.
But as we go through the conversation, if you have any
questions in particular for Albert, please throw them in
there. We'll, we'll either take them
live or we'll, we'll look at them at the end as well too.
So you mentioned, you know, the toolbox and it's adaptive and
it's learning, but let's step back a little bit.
(14:17):
Obviously, you know, agentic systems are super new for a lot
of people. So how much of what you're doing
today is also education, Albert?Like are you, you know, this is
super new for a lot of companiesand potential customers of yours
to allow kind of agents run and manage what they need to do in
(14:38):
this space. So how do you get across that
level of education that's required or you know, does your
demo do that for people, customers and they have that wow
moment? How does it work?
Yeah, that's a good question. So we've actually started off
with approaching partners who are already aware of well the
(15:00):
internal sort of of of the mechanics of what we are doing.
So that was pretty easy and likeit's usually like R&D
departments, right. So they have like a couple of
PGS themselves. So but but you are right.
They already have the pain pointand they know how complex this
is and therefore quite happy to see how you can help.
Yeah, absolutely. So that was the the case.
(15:21):
But of course we are there is a lot of education going on from
our side. So in fact, we, I mean, first of
all, right, there is a whole space of this agentic AI as they
call and like it. It's actually such a broad term
and like many people mean different things.
There is a lot on well the land side like which I mean
(15:43):
Transformers essentially just enable to compose this building
blocks of different a genetic systems.
And now they're like different protocols both from anthropic
and Google. And we, we are not in that
market. I, I would, I wouldn't say that
we are in a like a sort of a competitor.
(16:03):
We are rather complementing sideof, of, of that story.
And the complementation comes from from the fact is that we
are basically enabling or providing physics.
We are like providing physics for this identic system.
So right, as I said, like with the drone, we we provide some
(16:27):
physical properties. We kind of grounded in the
probability theory and we grounded in the physical world.
So in, in that way, of course, it's, it's, it's a new newish
approach, although it's been around for some time.
We do have collaborations with some institutes, well, which is
(16:49):
like active inference institute in I think their base and
states. So where we sometimes give some
tutorials on how we can create this type of authentic systems
as well as actually our open source part of our venture,
which is called ARIX Infer, right.
So it's arixinfer.com where is the a lot of educational
(17:10):
material. So like starting from very basic
stuff on probabilistic modelling, how to solve, how to
figure out that the coin you aretossing is fair or not fair.
So this type of things and goingall the way to how to build a
drone. So it's, it's really a granule,
I think approach to, well, to educate people on how to, yeah,
(17:33):
utilise the systems and utilise this approach is to build this
type of organic systems, yeah. OK, OK.
And so obviously we're on the Mongo DB podcast.
Where does Mongo DB fit into thetechnology stack, Albert that
Lazy Dynamics have? How are you using Mongo DB and
its capabilities? Yeah, I, I think, yeah, I, I
(17:54):
have actually a bit of a story with Mongo DB.
So I, I first, yeah, I, I first encountered it during my
bachelor's when I became interested in, well, actually
MapReduce algorithms. So I wanted to do, yeah,
something back in the day, simple with Twitter data.
So like let's say, finding popular, popular hashtags or
(18:15):
similar, like analytics and Twitter data, it includes really
diverse contact content types, right, that evolve over time.
And then I've, yeah, stumbled upon Mondo DB, which was
essentially allowing storing this document based structure,
right, like various data types, rigid schema constraints.
(18:35):
And yeah, This is why, yeah, I'mjust started starting to work
with that. And interestingly enough, I
think it's also a good lesson for for me.
Perhaps it is like why? Why you should read
documentation because well, somewhere along the lines of
Mongo DB docs, it was 2015, I actually think.
(18:57):
So they were, it was written that guys, if you, if you're
trying to use this MapReduce forthis particular problem, you are
doing it wrong. Like there is like some
aggregation functions that you can make use of.
So like just just don't, don't use my MapReduce for that.
So, yeah, so I've spent quite some time like which which were
gone in vain. Well, and at least I, I learned
Mongo DB also learned a bit of Python because the core calls in
(19:19):
Python. So, OK, it was good.
And then actually just eight years later, yeah, 80 years
later, I think so when we started to work in our company,
we realized that our engine which was working fine in
academic settings, but when you bring it to the industry, right,
it requires many additional elements, right?
(19:40):
Servers, backlogs, data safety measures.
So, and then, yeah, I was just like, OK, what, what, what,
what's now out there. And then I came across Mongo DBS
again, as a, as a, as a, an interesting solution, perhaps
because you can imagine like for, for drones is 1 type of
data. But when you are like doing
finance or something else, it's a different type of data.
(20:03):
So again, like rigid schema. I pitched this idea to our CTO
Admitry and yeah, here we are. So basically it's flexibility
and performance characteristics made it well ideal fit for our
dynamic data needs. OK, excellent.
So the yeah, the, the I can understand why, yeah, in the
(20:25):
with the Lego blocks that you'reoffering and the different areas
in which you can work a rigid, yeah, rigid structure, rigid
schema wouldn't work at all. And, and to me, it's, it's
interesting to see that pathway,you know, that kind of
eight-year gap is because obviously in developer relations
where I work, we're, we're, we're about enablement and
awareness and planting the seeds.
(20:46):
But as I keep telling my VPS here, it's a long time
potentially for that seed to grow.
And in this instance you had familiarity with it and and when
the calling came, you went, I think Margot Abe could be very
suitable for this. What in particular, I mean, are
you taking use of time series data?
Are you using a Mongo DB Atlas vector search and and what we
(21:09):
have there now what what parts of of our ecosystem that you are
using at the moment then Albert?Right, right.
So I will, I will go into the details of how we use it.
At the moment we are indeed working with.
So we, we've started to work with this leverage in time
series database because it's a critical dimension in our
(21:29):
system, right? So basically Mongo DB allows to
well improve our ability to process temporal data
efficiently. And this is just, it's not just
technical optimization. I think it's fundamental to how
our agents understand and respond to the world.
And I must say that we, we haven't utilized this
(21:51):
functionality fully yet. So it's still something we are
working on and like perhaps actually we will jump on it
pretty quick pretty soon. As for vector search, yes, so,
and this is this is something with which is really cool in in
Mongo DB. So one of the pilot projects
(22:14):
that we've been working on it actually well, it it's we had to
work with a lot of textual data.So it's actually textual data
spread in time also. So it's like you can imagine
there's like some text that yeah, we have at well it one
hour time steps. So and then of course, well, our
(22:38):
agent well to enable this sort of navigating through these
documents, well, we did have to provide it to some, yeah, some
vector searches like what traditionally done in The Reg,
right. And this type of approaches.
So it's basically a way for. So vector search was a way to
(23:00):
navigate this embedded vector oflike that is embedding
representation of of the textualfiles which was working pretty
smooth. So unfortunately I can't show
that demo because it's under NDAbut but I will show another one.
OK, perfect. And you mentioned some customers
and some types of solutions, youknow the robotics, the drones
(23:22):
etcetera. Are there other industries that
you know can benefit from lazy dynamics solution as well too?
Is it, is it super adaptable, super scalable for other areas
or who are who would you consider your primary industries
that you would go after in termsof business within Lazy
dynamics? Right, right.
So I, I, I think that the ideal customer is like the
(23:46):
organisations that are building agents or autonomous systems for
as I said, mission critical tasks so that face significant
uncertainty in complete data andneed to make plans despite the
noise. So, and in fact, like currently
we are on boarding Co development partners across, I
(24:07):
would say 4 primary sectors. And like we are basically in
this stage where we're just trying to figure out, OK, what's
what's the most lucrative at themoment.
But it's so one is autonomous systems, yeah, particularly like
drones or wheeled robotics, thenlogistics, which is kind of
interconnected with that. And yeah, predictive maintenance
(24:29):
is basically to figure out, OK, when your reporting system or
when your, your system will failand like what, what is the
uncertainty around that that failure?
And yeah, there were quite some interest from finance industry,
but yeah, we are, yeah, not well, we'll see.
So and, and but but this interesting this as I said,
they, they share the common need.
(24:50):
So they require systems that canadapt to changing conditions
while maintaining this reliability.
Yeah. So that's what connects them
all. Which is, you know, it's, it's
nice to have that super broad audience.
And I think going back to the earlier conversation about
agentic systems, I think it's really early days to figure out
(25:12):
how and where they might be applied as well too.
I think we've all seen the clever demos, you know, but you
know, yeah, agentic system booking my holiday for me within
my budget in the and getting my travel itinerary.
Yeah, nice demo, right. But I'm really looking in all of
these kind of demos under the hood at, you know, where's the
(25:34):
production application, where's the scalable application, you
know, outside of these, you know, show and tell demos.
And I think you guys are obviously addressing that area
very much. So you seem to be at the at the
cutting edge looking ahead beyond this maybe what's the
longer term vision for laser dynamics at this time, Albert or
(25:55):
you know, kind of what else, what other areas could you get
into perhaps too? I'm conscious that we'll, we'll
show some demos. So I'm trying to just tee up the
understanding of how this could be applied perhaps?
Sure. I I, I think that just just to
linger a bit on, on the example you've provided with this, you
know, finding the, the trade, like within the budget.
(26:16):
I think it's, it's a great application actually of well,
it's a genetic system based on, you know, lambs, right?
Or like VLA, essentially Transformers.
But I wonder actually, are thereresidents where people actually
just drafted with their like a whole heart like to any of those
(26:38):
right. Like, yeah, I'm, I'm not.
So I I've seen quite some peopleactually well on boarding this
SDA SDRS, right, the sales development robots and yeah,
like a month my. Sales development robots as SDRS
like in Mongo DB, internally we have SDRS, but they're sales
(26:58):
development representatives. We haven't replaced them with
robots yet, have you? So you're seeing this?
Well, I, I do know that there's some people actually, they've
tried and they've, they've hatedit as well, right?
So, and the reason is the following again, right.
So you're basically, it's, it's not a compiloting system.
It's like a really autopilot where you basically bet your
(27:21):
business, you bet your reputation on, on, on the LLM
that can hallucinate. And and I said right, like for
this, for the spam, I can be perfect.
I would like for creativity, yes, but would you bet your
well, let's say life your business on that?
And I'm not so sure so and I don't know how our approach can
(27:42):
address that yet. I, I, I really like I'm we're
really in the scope of more likephysical world, like on, on,
yes, something like really tangible, some signals that we,
we, we know what is the meaning of them.
So and well, I'm not saying thatthis identic system that are all
now an owned well, what they're,which are now, which everyone is
(28:04):
talking right now about. It's, I think it's, it's, it's
it's a great deck for sure, but it's a bit explainable kind of
like it's a bit like you can't really figure out and trace back
OK, what what really happened there like, or if you if you can
do it, it can be really costly. Like with our system, it's well
usually like very much explainable.
(28:26):
So we actually do know all the time what went wrong in the
system. But because it's model based
approach, right? So it's, it's also does it I
know also to sell it is like this is the silver bullet for
identic system. No, it, it solves particular
problems within identic systems,But well, we'll see.
It's it's still a long way. And when you ask like about the
(28:47):
vision, I think the stretch goalis for sure to empower this
planning abilities of hygienic systems.
So essentially to become a brainfor well, autonomous systems or
let's say robotic applications. So I really want to envision the
world where autonomous system, they can navigate complex in
certain environments and yeah, enable this that the ability
(29:13):
decision making capabilities, yeah, that were previously been
impossible. So we can actually trust them,
yeah. OK.
OK. Well, I'm sure with that team
that you've collected around youin Laser Dynamics, I think very
much look forward to how things shape out.
Adita, thank you very much for the comments as well.
Appreciate that if there's any other comments, our questions
(29:36):
actually for Albert as we go through, please do let us know
in the comments and post them aswell too.
So I think like always, I try toget the guests to showcase
something and show us a little bit of a demo.
So I think given the high level conversation that we've had now,
Albert, I think it's probably about time to to showcase some
(29:56):
examples and a practical exampleand showing us kind of what
we've been talking about for thelast 30 minutes or so.
Right, right. OK, so let's start about that.
I'll just make sure that everything runs.
Yep, I think she'd be good. OK, so screen.
(30:17):
Yes, screen. OK, do you see my screen now?
Yes, that's it. It's, it's coming through clear.
We see lots of lots of super interesting stuff.
Albert here. What is it?
What are we? Looking at yeah, OK, OK, so so
this is actually a digital simulation of of the drone
environment. So like think of drone as an
(30:38):
agent and there are like 3 different tasks and scenarios
for for this drone. I'll start with what it's doing
and then how we do it actually. So so the first task as you see
here, so we have a drone which is basically continuously tasked
with going from 1:00 point in the well in this 3 dimensional
(31:04):
space to another one. The thing about this drone is
that it doesn't know anything about the obstacles that it will
see in front of it. So the idea is that OK, so the
drone is tasked. So you say, OK, drone, you you
should go there and then as longas it starts moving, it
encounters different objects andit needs to adapt in real time
(31:27):
to to avoid those. So as you can imagine, drone
doesn't have a luxury to stop. It's to pause.
It's thinking, you know, like, what would I do next?
So I mean, doesn't work like that.
So you really need to keep reacting on whatever
environmental changes occur. So here we're just basically
(31:47):
showing how drone sees its distances, like towards the
Yeah. So another yeah, the kind of
small example is that the drone here is essentially was not, not
it knows that it needs to to do this sort of simple pattern, but
what it doesn't know that it's it's raining outside.
So obviously when it's raining, this environmental condition
(32:10):
changes for the drone, so it needs to adapt.
OK. And that's what we what we show
here. So OK, there is sort of an
adaptation going on like as, as it's moving, so it tries to
learn also OK that it now needs to operate with its actions or
controls differently. So what does it have to do all
(32:31):
with Mongo DB? So and perhaps before going
there, so there is actually another front end.
So it's I mean, it's kind of real time demo of right the
dumping from so the the point where it needs to go generates
randomly and like as soon as it pops up, the drone kind of
figures out this action space where how it needs to control
(32:52):
its rotors to actually reach there.
So in that you can see it's all kind of happens real time.
So we can have the statistics over the engine forces.
There's also a bit of code how how it works.
But let me get first into the yeah, what what Ericsson, what
the Mongo DB has anything to do with that?
(33:13):
So so this is the demo that I just showed right where the
drone is just to call this goal.And there are three components
here. So the third very first
component is this front end is like basically your simulation
or in the real world, it would be just a real draw.
I reconfirm it's a, it's a server here which is running our
(33:38):
engine. Like you can think of it as a,
as a brain kind of that decides what actions sent to the real
world, the right. So it's kind of the inference
engine as we call it. And that's the open source
project that you gave me the link to earlier as well too,
This one here, OK. Yes, yes, yes.
So this open source, I can play,I can talk a bit on what's not
(34:00):
open source, but later on so, and then there is Mongo DB.
So why all matters? Because, well, all these things
need to communicate to each other and they need to
communicate each other pretty fast, right?
Because, well, there is of course the server and the front
end or like the real world wherethe communication needs to
happen real time. And we would also prefer to have
(34:22):
Mongo DB communicating to the server also as real time as
possible. So you don't want to have this
slacks and of course, and that'sessentially what's going on.
So there are like 2 communication pipelines.
So first is like from the serverto Mongo DB and what it logs is
current states data logs, right?What was happening with thinking
(34:44):
process, right? So how, how did we end up to use
that particular sequence of actions in order to get to this
place? We also store previous states,
the parameters of the drones, right, Because it could happen
that we don't have the full specification of of the drone.
For instance, we don't know it'sinertia.
So it's something we need to learn and we want to track over
(35:06):
time how the drone was learning that.
So this is very important and also for basically analysing it.
Well, if all of a sudden this drone flew to someone's else
window, right, we want to know what what actually went wrong
there and here here. What's important also is that
sometimes why this rigid schema comes into play, sometimes
(35:29):
sensors fail right? So it's something that the real
world can surprise you with, right?
So this can be like a a failure in the sensor in the road or
anything. And then it creates some null
values or like missing values. And that that's something we
also want to to log right away. And another pipeline is is from
(35:51):
server to the front end. So yeah, it's I see here it's
yeah, 3030 events per seconds per second.
And it's multi dimensional observation, right?
So we actually observe the droneobserves different well,
different data. So well, you can imagine like
it's from camera, it can be one type of data, right?
(36:14):
There is also, well, let's say the position from IMU sensors.
So it's really different than well, kind of scars data we are
working with. And yeah, so here what's going
on is like real time state estimating, prediction and
quality generation with Mongo DBas a back end database and it
(36:35):
works pretty well so. Good, good.
I'm glad to hear. Yeah.
And then yeah, just three things, what we are using Mongo
DB for as well. So we've talked about different
applications, some of which I'm able to show, but so for real
time embedded systems, mod DB turned out to be really good,
(36:57):
like for a thorough pool drive, minimum latency, fast date
updates. We also, as I said, like we're
working with unstructured data or like badly structured data
when you're working with texts usually like that.
So, yeah, there was a good use of Mongo DB vector search that's
(37:21):
also sort of a way for our agentto navigate the the environment,
OK to find, OK, what are the documents are relevant and well
predictive maintenance again, very similar, very similar to
real time embedded system. So we want to identify these
historical partners and react assoon as possible.
(37:42):
So yeah, as just to conclude, right, different patterns and
same database. So that's that's what we like
nonvidible for nonvidible before.
Excellent. I couldn't have said it better
myself. So no, I I'm glad.
I'm glad that you went through that.
That's that's perfect. We did get a question in from
from Janani and I hope I pronounced that correct.
(38:03):
And she's asking what kinds of path planning algorithms or
approaches do you use when the agents are involved and when the
obstacles in the environment areunknown?
Is that on yourselves and lazy dynamics or is that on the
client side when they use lazy dynamics?
No, that's a great question. Thanks a lot, Gianni.
(38:25):
So The thing is like, I would have to go into a bit of the
details of Eric's and Fer. So it's on our side, first of
all. So that's something we figure
out for for the for the client or like for the user,
essentially. So the way work and I really
advise John, I need to go into the open source or like reach us
(38:48):
out on the on meet hub. We can elaborate on that.
So first of all, we use we we caused this planning problem as
sorry for for for the slang, butbut as as a Bayesian entrance
problem. So we basically say as a the
planning problem can be reframedas a probabilistic entrance.
(39:13):
So think of it this way, right? So when you do planning, you
will basically say to your agentor the drone that you want to be
seen where in some time now you can like you, you can think it's
kind of inverse problem, right? So you're trying to figure out,
OK, what is the optimal, optimalpath?
However, why you can't solve it,for instance, with a star, which
(39:37):
is like a very common algorithm for path planning is, is that
the environment which we're in is dynamic.
There are many uncertainties around your sensors.
There are many uncertainties about other obstacles.
So we basically say to the agentthat it needs to dream about
this path 1st and usually when the when the agent dream about
(39:57):
this path and like that, that's kind of what we do in the
server. Of course it, it has very
optimistic agenda, right? So it thinks that well, there
will be no obstacles or maybe there will be yeah, well, there
will be no wind or something. So it tries to draw this very
simplistic path. Well, which goes ahead with this
Oh, with accordance to least acting principle, I will not go
(40:21):
into there, but sorry, sorry to mention.
And then? When the drone starts to move
according to its planned agenda,which we provide, which we, we,
we gave you this plan which you can execute, what starts to
happen is that well, there are like, well, some obstacles
appear or some failure secure and then it immediately tries to
(40:41):
well circumvented by replanning.And this replanning is usually,
well, it's it's a very expensivetask to do.
So there is a whole way of doingit in our way.
So we're we kind of developed this a smart way of planning,
let's put it this way. So we actually managed to replan
in real time. As for the yeah, As for the
(41:03):
obstacles, of course the the, the drone needs to sense
something. So there should be some sort of
a sensor that can say that the obstacles otherwise, well, we,
we are in trouble, right? It's like, yes, so, so there
needs to be a camera or like some, some lighter or whatever
(41:26):
that kind of can can tell you OK, how far you are from from
the office. OK.
And forgive me if you've alreadytalked on this, but the that re
planning obviously that's all automatic.
Is there some feedback loop backto, you know, back to the
customer who's using this that that has occurred?
Or is it go and do the job, takeinto account of the paths, the
(41:48):
obstacles, the environmental conditions and finish the job.
And yeah, so, OK, so this is where open source differs from
progression. So in in open source version,
you got to figure it out yourself how you're going to
utilize all this replanning and planning and executing on the
plan together. And I mean, we're, we're helping
(42:13):
with that, right, as, as much aswe can.
Like on our, well, we didn't have examples or GitHub
discussions, but it, it's, we also recognize that it's hard
and that's actually where the value position of laser.
And so we automate that. So we basically tell you, OK,
forget about all these worries, how to sort of create these
(42:35):
loops or feedback loop as you just mentioned.
And we will just do that for you.
So all you would have to do is to pre specify essentially to
call just three functions, whichis plan, act and and learn.
So and and all the rest. We are we are doing this magic
for you. So yeah.
OK. Oh, brilliant.
OK. So that's the pathway from open
(42:57):
source to to revenue as far as lazy dynamics are concerned then
so. Right.
OK, we'll look it makes a ton ofsense.
It's a well worn pathway for forservices.
Mongo DB is open source essentially, but if you want all
of the the enterprise features or if you want to use Atlas in
the cloud, it's it's our software as a service platform.
So that makes sense as well too.So it's a it's a well worn
(43:21):
pathway in Ireland, where I'm based, we have a company called
Mana Aero that does drone deliveries and you know, it's it
seems amazing. It seems fabulous what they do.
You know, it's, it's for low value items, you know, copies
and fast food and takeaways and things like that too.
But I, I can imagine, given whatyou've discussed, the
(43:41):
complexities that they must haveto overcome the particularly the
environmental stuff. As I said to you when we were
jumped on the call, it's, it's areally bad weather day here in
Ireland. I can imagine that that affects
the flights and, and paths of those type of services.
Are you seeing those services, they always seem to be kind of
trial. Are you seeing those
(44:02):
commercially used? I know there's been some nice
case studies in in, you know, Africa and elsewhere for medical
supplies deliveries as well too.Is that any of your customers
today? So As for the robotics right
concerned we're we've just started to onboard partners.
(44:22):
So I can't really disclose them.So we are, but they're like fast
like forward thinking customers who who recognize that and then
we we, we are providing this sort of a solver for their drone
dynamics. So and I, I do, I do believe
that in the future, that's how the yeah, mass delivery will
(44:47):
operate. But there are quite, quite many
issues that needs to be addressed right.
So well, a lot of liability issues with that then there is
obviously this well sound noise,right, this this pollution,
right, this pollution. So this is something that yeah,
(45:10):
I think this drone deliveries will will face.
I know there are companies whichare actually trying to well
develop the drone such that theybasically are almost silent.
So but that's, I think also a few years away.
I, I remember when I was doing my masters in, in, in Japan, I
think Rakuten was also just about to launch their drone
(45:33):
delivery program in. I've been there again.
I've been there again, like 3 weeks ago, it didn't
materialise. So yeah, I think it's hard.
It's hard. And as you said, there's a lot
of uncertainty around those deliveries and like how also
people approach this drones, right?
So it's there, there are many issues that if you look at like
(45:57):
from kind of surface perspective, seems like wow,
it's, it's easy stuff like you figure out.
But yeah, the devil is in the details.
Yeah, very much I think, yeah. As you said, there's a lot of
these companies just about to launch commercially and and it's
always just about to launch. You mentioned I think at the
intro as well, probably the moreunderstandable scenario of, you
(46:20):
know, warehouse and there's robots and there's automated
systems and warehouse that do things.
We see from time to time the large retailers, you know,
automated warehouse, etcetera. Obviously the drone example is
one that catches the attention, but I would imagine the scale of
the warehousing robots automation is something that's
an enormous challenge to tackle as well too.
(46:40):
Oh yeah, yeah, I. Thought Lazy Dynamics can offer.
Yeah, I think, I think I can also sure share that.
I think we'd have time, right? So.
Yeah. That's.
OK. Janani came back and she said,
you know, she learned some probabilistic path planning in
class and why she asked the question.
So she she thanks you for your insights there, Albert.
(47:03):
Thank you. OK for the question.
Yeah. Thanks a lot.
Yeah, yeah, a lot of these questions for from people who
already know the answer, right. So, but yeah, OK, good, good.
Now I have yeah, I just invite her to to check us out.
So I think she would have it. Interesting.
Yeah. So, yeah, there's a small, just
another example. I think you just mentioned this
(47:24):
warehouse. And I think this is also perfect
kind of situation where we, our solution is very handy.
So here you have this kind of, yeah, like a bird's eye view of
a warehouse. So you have these four agents
that are tasked to go from one location to another and
(47:48):
locations are represented as stars.
This small agent, they also havesort of this, you know, this
death circle around that, which is basically their well, their
sensor like where they can locate the obstacle and the task
by itself. If it was just a static
environment, it's not that hard.Like it's really not that hard.
(48:11):
I mean, especially well, it's it's static environment only
robots are there. But imagine it's not only agents
like I mean robots, but they're also humans.
And and not not to say that onlyhumans make mistakes like also a
genetic systems like robots alsocan well block the aisle or do
something nesting like that willdisrupt the the plant.
(48:34):
But here we actually do is a very, let's say hostile
environment, like basically an environment that changes all the
time. So when the agent tries to
execute its idealistic plan, right, like is it is it tries to
sort of find to to dream about how it would go that the the
(48:55):
environment always pushes back. It doesn't let the agents to
materialize. So they actually need to
navigate seamlessly. OK, how do I avoid this?
How do I avoid that as they go? Because yeah, again, we we can't
stop some of the processes whichare ongoing.
So the task was here to, OK, we define the requirements for the
agent. That's actually example.
(49:16):
I think also now we're open source, which is I think like 15
lines of fault just to define this model.
And the the task is the following.
So respect the speed, respect the distances and go from one
location to another. And then, yeah, the agent, they
figure out how to do that. OK, OK.
Yeah, an incredibly, I suppose under the hood, we don't
(49:37):
generally think of these things to this level of detail.
But obviously, of course, and, and as you mentioned the the
driving example originally, you know the environment is ever
changing and and you just can't,as you say, I love the term that
you use dream about the perfect path, but that never probably
ever occurs, right? Yeah, absolutely.
(49:59):
I mean like, like, like for us humans, right?
I mean, just saying you were like, do you think that
something will occur in like andit just happens in a completely
different way? And nevertheless, yeah, we
managed to adapt. So I mean, I, I haven't
mentioned that, but like actually our professor, right,
like he was, he's been always fascinated about how biological
systems adapt. And like many of things that
(50:22):
we've implemented also actually tried to draw some of the
inspiration from what he's been thinking about.
So yeah, how how we human adapt reactively as the environmental
environment changes. So that's that's an interesting
topic for sure. Definitely, definitely.
Oh, this has been fascinating, Isuppose.
(50:42):
And look, and obviously that thefact that you've done a
doctorate etcetera as well too. You're an avid learner, Albert,
but you know, this space as you,you know, look for those that
who don't know anything about this space, you think AI is only
2 1/2 years old, right? For anyone who does, it's been
machine learning and artificial intelligence for, you know,
many, many decades. How do you keep up to date with
(51:05):
the changes in this space and and where do you go to to learn
and keep ahead of things for Lazy Dynamics as the CEO?
Yeah, that's a great question, Shane.
I I should say that I think thatwhat, what's really important is
to build the community around you that led to different things
(51:26):
specifically. Yeah, I have friends who are
thinking that, yeah, Transformers is this best thing
since sliced bread rice. We're really like very like
venturing into that and I respect that I, I think that
even though, yeah, I, I disagreewith them, I'm trying to keep up
with that. So I, I think that for us,
(51:46):
what's important is to really build up the community around
things which are complementary or competitive to us.
So just to stay, to stay kind ofrelevant to see what's going on.
I think there's also a lot to learn from these things, right?
So you can, you can have all this other agentic systems,
(52:07):
right? There's a lot to learn there.
So from realization perspectivesand so forth.
And so this is basically the waywe, we stay relevant.
And I think what's most important is for, for our
companies that we, we are reallyventuring in.
So what we do, it's very important to believe when what
you are doing right. So it's of course not this
(52:29):
pretty shared kind of stuff, butthat, that that's the reality.
So we, we do think that our approach is, is the best one, is
the right one. And like for specific problems,
it will be, well, it, it will draw others obsolete.
And I think that this is very, this is very important.
So yeah, I guess these two things, community and for
(52:52):
believing in what you're doing, yeah.
Yeah, I love the community aspect of that.
And and I suppose obviously you're, you're getting feedback
from the early customers, the early adopters of Lazy Dynamics
as well too. And you're enabling them and and
growing with them as well, whichis, which is amazing to see.
It is a yeah, it's, it's a fascinating area that you've
(53:13):
you've gone into. And I think obviously timing is
everything too, Albert, Right. So you were, you were doing work
in these spaces, you know, throughout for a vast number of
years obviously and and then sawthe commercial prospect of what
you've been doing and gathered that team around you.
Yeah, that's right. That's right.
(53:33):
I think, yeah, I shouldn't have mentioned the partners and
customers because this area the well, the real asset who, who
kind of makes this reality checkfor the product, right?
So because again, like as, as founders from academia and, and
they see it a lot, right? Like everyone who is trying to
launch out of the lab, they're trying to basically they have an
(53:56):
amazing technology and they're like trying to find the fit for
it. And, and, and we were like, just
like that. So with that, yeah, we have this
amazing technology, but I mean, truth to be told, no one cares,
right? So it's about the the right
problem that your technology cansolve.
So, and this is hard and like assoon as you realize that you
(54:17):
find your customers and you actually get very valuable,
valuable feedback actually that shapes, shapes your products,
shapes your technology in the right way.
Yeah, no, it's great. And I suppose given the
background that you all have with all your PhDs behind you
and all the different areas, I, I used to work with a lot of
(54:39):
startups And what we find a lot of the time as startups are
building products, but you know,they're, they're afraid to go
out there into the open and get customers because it's, you
know, it's, it's a real world validation of their system.
Whereas you seem to be very clear on, you know how Lazy
Dynamic can help these customersalready and the environments and
(54:59):
the applications to which they can be beneficial, right?
Right, right. Yeah, That's that's, that's a
great conclusion. Yeah, I think that's.
Yeah, well, like for for a company 18 months or so old, I
think you've come on in in leapsand bounds Albert.
And it's, it's great to Share your story and the the journey
that you've had and give us kindof the glimpse of the insights
(55:22):
into kind of the innovative worlds that you occupy in Lazy
dynamics before we wrap up. And and I didn't see any other
questions come in. So obviously, as I said to you
on our prep, usually that means you've been very clear in your
mission and what you set out to explain here.
So that's that's great. Any last words for our audience,
maybe from the point of view of,you know, going out there and
(55:45):
found in a company or, you know,what's happening in the identic
world or how to keep up or how to learn?
Yeah, I think we've touched uponall of these topics.
I think that, well, there's justone thing I would, I would love
to to ask people for, I guess isjust to check out Oryxin Fer,
(56:05):
check out Laser Dynamics. Oryxin Fer is completely open
source and I'm, I'm sure that everyone, every expert in their
own field would be able to relate to some of the problems
we highlight on our open source website.
And well, just reach out for, for, yeah, for anything like it
(56:28):
just be the open source questionor like the possibility for
potential project together. So we're open for any sort of
collaboration. Yeah, that's that's basically
it. Excellent.
Well, listen, I think that's a good call out to for to wrap up
our our podcast here, Albert, it's been a pleasure having you
(56:50):
on board and sharing. I will certainly keep an eye on
what lazy dynamics are doing. And you know, whenever the next
stage comes around or whenever the next demos that you have,
I'd love to get you back on the show as well to to explain.
But I think for now, obviously, as you said, go to or X infer
and also as you see on the screen there at the moment, Lazy
(57:12):
dynamics.com to learn more. But for now, for me, Shane
McAllister, thank you so much, Albert.
It's been a pleasure to have youon board and all the best with
you and the team at Lazy Dynamics for the future.
We wish you every success. Yeah.
Thanks, Shane. Thanks.
Excellent. And for all our viewers, thank
you very much for joining us anddo keep up.
(57:34):
As I said in the introductions, if you like and subscribe to the
Mongo DB YouTube channel, you'llget alerts for similar live
streams such as this. And also follow us on LinkedIn
to get similar event notifications there too.
I stream most Tuesdays at roughly this time with lots of
amazing guests such as Albert. So do join in.
But from both of us now, thank you all for tuning in.
(57:56):
It's been a pleasure to have youand Albert.
Thank you once again, it's been great.
Take care everyone, good luck. Bye bye.
Bye.
Hello, welcome to MongoDB Podcast Live.
I'm Shane McAllister. It's great to have you.
And today we're diving into the fascinating intersection of
robotics, machine learning and real world applications with the
true innovator. And to elaborate on the amazing
title of turning chaos into opportunity, please join me in
(00:20):
welcoming Albert, the CEO of Lazy Dynamics, a company
revolutionizing how systems react and adapt in dynamic
environments. Albert, you're very welcome to
the Mongo to Be podcast. Hi, Shane.
Yeah, thanks a lot for this introduction.
Happy to be here. Not at all.
It's great to have you. Thank you for joining.
(00:40):
I always start and I know we dida prep call on this.
I always start with our guest totell us a little bit about their
career path to date and how theygot here and in our
conversation, Albert, you know, that was incredibly interesting.
So I know we don't have all the time in the world, but maybe
perhaps give us a potted historyof of your background to date
(01:02):
and and how you ended up being the CEO of Lazy Dynamics.
Sure, sure. Yeah.
I'll try to compress it as much as possible.
But yeah, so my background, so basically I hailed from Far East
of Russia, a place where far from where I am based right now,
which is in the Netherlands. I completed my bachelor's degree
in the western part of Russia, like at Saint Petersburg State
(01:25):
University, where I was studyingapplied mathematics and control.
So upon graduation of my bachelor's, I, I think I was a
bit likely to get a scholarship for a two years master program
in, in Japan, in Kyoto, in computer science, where I dove
into self driving cars and driving aids research.
(01:47):
But it was odd. Well, it, it's still getting it,
it it's still odd. Or like, let's say it's WAVY,
right? It's like always five years
away. Back in 2016, it was quite hot
was coming in two years and and yeah, so I would I embark on
that research where we were developing essentially an an
assistant for elderly drivers. So that could predict their
(02:09):
intention to emergency brake before they even hit the the
braking pedal. And the research, I must say,
like progressed well, we even secured a patent.
But when we tested it with real drivers, my friends, we hit
quite a roadblock. So they basically hated the, the
system. They, they really didn't like
(02:30):
it, to put it mildly. And well, and the reason was
that it generated too many falsepositive.
So it was trained on the so-called average driver, right?
And like, well, now it's retrospectively it seems
obvious, but well, back in the days, like it wasn't.
So the, the problem with averagedriver, it doesn't exist.
(02:52):
So, so and then it led me to explore like adaptive systems,
adaptive technologies, how couldwe kind of leverage that prior
knowledge from many drivers thatwe've well that we've researched
right to, but then to fine tune itself in real time.
So how can we kind of have a system that can fine tune itself
(03:15):
in real time to the actual person behind the wheel.
And I think this eventually led me or guided to my PhD working
on adaptive hearing aids. And it, it may seem that it's
completely orthogonal problem, right?
Like where is like the driving cars or like where is the
hearing aid patients, right. And in a way there are different
(03:39):
problems, but what connects them?
There is no average hearing aid patient.
There is no average driver. And for both tasks we actually
require to have a system, or nowas they call it, an agent that
can adapt into the ever changingworld.
The acoustic environment changes, right?
(03:59):
They're all changes. And then also there's kind of a
sense of the personalisation, like we want to personalize the
hearing aid algorithm to basically work for a particular
human driver. Sorry, for a hearing aid
patient. And yeah, that's basically what
I've, I've been working on during my PhD.
And yeah, after in four years we've basically developed the
(04:23):
general toolbox, a solution as we call it, an engine that
enables this type of adaptive learning.
And well, about a year ago we spun out from our laboratory and
I guess, well, everyone in my team, Co founding team took nice
roles and like I was assigned asa CEOI would say, well, so
(04:47):
that's, that's kind of how it worked out.
Yeah. I love that assigned as ACEO.
So let's talk a little bit aboutthat.
I suppose maybe the team itself,the diverse academic background
that the team has that you've assembled in Lazy Dynamics and I
suppose in in the pathway of talking about that team talk a
(05:09):
bit about the driving force of the the spin out from from the
university then as well. Albert Right, right.
So I, I think the, the, the coreteam consists of five PhDs, so
each bringing, bringing criticalexpertise like we have our
mathematics guy, our coding expert, an engineer, and of
(05:31):
course the research lead. And I guess, and there is me
like serving as a as a CEO. So precisely because I'm not
excellent at any of those other specialized roles.
So yeah, but I'm. Sure, you're very much
underestimating yourself there now.
Yeah, that's what they say. But, but I, I think, I think I'm
definitely good at bringing it all together.
So that's, that's basically, I guess a connection point for all
(05:54):
that. But that, that background is
rather diverse, I would say. So we actually all came from
different masters and, and the bachelor's.
So some people were like doing mathematics, some people were
studying neuroscience. And then after all, it sort of
started to converge into one solid idea on creating this sort
(06:17):
of engine or software, software toolbox that merges many
different ideas to create, well,one solid framework to help to
enable, as we call it like now, adaptive intelligence.
OK. OK.
So from a timeline point of view, that team which as you
mentioned a blend of maths and engineering and coders etcetera
(06:40):
and the scientific knowledge behind it all that that was
about a year or more ago then was it Albert?
Yeah, I think so. So we we were incorporated 1 1/2
year ago, but we truly started to work I think like 2024.
And the reason for that was simple.
So that when we established the company, they were some of our
(07:00):
Co founders who are still doing PhDs like actually like they
were very close to dance which was not an option to do a drop
out. So which normally founders do.
But in our case, it was like just three more, three more
months that of course extended to half a year.
But nevertheless they finished now.
So now we're all business basically.
OK, wow, wow. So look in your own case, a big
(07:24):
a huge background there and variety.
And as you said, there are similarities between, you know,
the self driving and the hearingaid.
But tell us a little bit about what lazy dynamics then is is
focusing on. Yeah, I I think the best way
actually to go about it is like to to well to go a bit backward,
(07:44):
it's actually like so to to to start with a bit of well
observation about this kind of founding story.
So, and the interesting observation is actually about
this current AI systems. So everyone is now talking about
agents, about entities essentially that perceive their
environment, take autonomous actions to achieve certain
(08:05):
goals. And the problem we identified at
Lazy Dynamics and we kept identifying throughout our PhDs
is that most authentic systems that that are based on for
instance, LLMS which are fantastic technology with great
implementation and content creation outreach campaigns.
(08:26):
However, when you when you try to build, let's let's say agents
for mission critical sectors, say autonomous systems, smart
cities, finance or hearing aids.OK.
The partners we are working on working with, they aren't eager
to bet lol. Based agents that can
hallucinate and struggle to adapt quickly enough.
(08:49):
And this is kind of exactly whatwe provide with our engine.
So identic systems based on probabilistic modelling.
I will not go into much details,not to this conversation, but
they basically we try to make agents that adhere to racial
reasoning and adapt rapidly to changing conditions.
(09:10):
And just to give you like a really dangerous example, I
think so you can imagine, let's say like an ambulance rushing to
a hospital. Yeah.
OK. And an agent which is inside
this emergency car, right? So it needs to find the fastest
route. So it needs to continuously like
(09:30):
process real time traffic, day data, weather conditions,
emergency AOV, equal locations to dynamically update this
optimal path. So and so unlike the traditional
navigation systems, our agent, it adapts its decisions based on
the changing probabilities. So say you are in the junction
(09:51):
and like now you can take left or right and the system tells
you, OK, if you go left, it willbe 6 minutes, if you go right,
it will be 5 minutes. However, I am very uncertain
about if you actually check turnright.
Yeah, this 5 minutes can actually turn to be 15 because
there are like some conditions there.
So the agent actually tries to proactively suggest alternative
(10:15):
by considering this certainty around certain decisions that
it's going to make. So yeah.
OK, OK. I think that's a good real world
example, but I suppose you're how are you, you know, what's
the heart of the system there that you're building?
You mentioned the toolbox earlier as well too, you know,
(10:38):
how does it and I understand that example of an ambulance and
there's a driver hopefully in the ambulance.
We haven't got to that stage yetwhere we don't have the driver.
But obviously you're looking towards, you know, drones and
driverless cars and other thingsthat are making these decisions
on their own, right? So you the toolbox you're saying
that laser dynamics have built. How is?
(10:59):
How is that addressing that bigger problem right?
So the toolbox is called Riggs and FUR and it's basically an
engine that allows to build agents.
So what we do is essentially we provide tooling for users to
(11:21):
create a genetic systems by using some Lego blocks as you
would like adaptive Lego blocks that we call probability
distributions or like let's say functions and allow to well
essentially describe the environment where the agent
lives, right? So be it.
(11:42):
If we are talking about the drone, we describe through some
high level abstraction functions, OK, how's the
dynamics of the drone changes like say if we if I'm here at
the point XYZ and I want to go to XYZ TT1, right?
So how, how this dynamics work? And The thing is that some
(12:07):
people they know these dynamics already, so they can just
provide those. But sometimes it's not really
known like there are like some robotic system for which you
might not have the full description of this of the
internal dynamics or of the model of the world.
So we basically create this additional building blocks which
(12:27):
also encapsulate the parameters of this unknown functions.
And the idea here is that once the drone goes out in the world
to solve the task and the task, we can mean like like a
surveillance or or we can think of going just from one from
point A to point B. The drone needs to figure out
(12:48):
OK, what type of the actions or like controls its needs to apply
in order to reach the goal. At first it it might have to
well to learn a bit. It's dynamic, so it needs to do
exploration. So essentially what we do is
like what our SDK allows to to create is well, describe the
model of the world for the agentand then run the inference.
(13:12):
So basically then you press the button once you've specified
your agentic system, your goal. So once you run our SDK, it
allows, well, it provides you with the solution, sort of a
solver how to reach that certaingoal in the environment you
describe. OK, fascinating.
And I know we're going to see a demo of some of this later on
(13:35):
towards the end of our conversation.
So for our viewers joining in now with Albert and me chatting
away about what's very complex, you are going to get an
illustration of how this works and what's it entails as well
too. And, and for anybody joining the
live stream, if you've got any questions as we go through, you
may not have missed seen the beginning where I was saying and
(13:56):
some people have shouted out said who they are.
Do let us know where you're joining in from as well too.
But as we go through the conversation, if you have any
questions in particular for Albert, please throw them in
there. We'll, we'll either take them
live or we'll, we'll look at them at the end as well too.
So you mentioned, you know, the toolbox and it's adaptive and
it's learning, but let's step back a little bit.
(14:17):
Obviously, you know, agentic systems are super new for a lot
of people. So how much of what you're doing
today is also education, Albert?Like are you, you know, this is
super new for a lot of companiesand potential customers of yours
to allow kind of agents run and manage what they need to do in
(14:38):
this space. So how do you get across that
level of education that's required or you know, does your
demo do that for people, customers and they have that wow
moment? How does it work?
Yeah, that's a good question. So we've actually started off
with approaching partners who are already aware of well the
(15:00):
internal sort of of of the mechanics of what we are doing.
So that was pretty easy and likeit's usually like R&D
departments, right. So they have like a couple of
PGS themselves. So but but you are right.
They already have the pain pointand they know how complex this
is and therefore quite happy to see how you can help.
Yeah, absolutely. So that was the the case.
(15:21):
But of course we are there is a lot of education going on from
our side. So in fact, we, I mean, first of
all, right, there is a whole space of this agentic AI as they
call and like it. It's actually such a broad term
and like many people mean different things.
There is a lot on well the land side like which I mean
(15:43):
Transformers essentially just enable to compose this building
blocks of different a genetic systems.
And now they're like different protocols both from anthropic
and Google. And we, we are not in that
market. I, I would, I wouldn't say that
we are in a like a sort of a competitor.
(16:03):
We are rather complementing sideof, of, of that story.
And the complementation comes from from the fact is that we
are basically enabling or providing physics.
We are like providing physics for this identic system.
So right, as I said, like with the drone, we we provide some
(16:27):
physical properties. We kind of grounded in the
probability theory and we grounded in the physical world.
So in, in that way, of course, it's, it's, it's a new newish
approach, although it's been around for some time.
We do have collaborations with some institutes, well, which is
(16:49):
like active inference institute in I think their base and
states. So where we sometimes give some
tutorials on how we can create this type of authentic systems
as well as actually our open source part of our venture,
which is called ARIX Infer, right.
So it's arixinfer.com where is the a lot of educational
(17:10):
material. So like starting from very basic
stuff on probabilistic modelling, how to solve, how to
figure out that the coin you aretossing is fair or not fair.
So this type of things and goingall the way to how to build a
drone. So it's, it's really a granule,
I think approach to, well, to educate people on how to, yeah,
(17:33):
utilise the systems and utilise this approach is to build this
type of organic systems, yeah. OK, OK.
And so obviously we're on the Mongo DB podcast.
Where does Mongo DB fit into thetechnology stack, Albert that
Lazy Dynamics have? How are you using Mongo DB and
its capabilities? Yeah, I, I think, yeah, I, I
(17:54):
have actually a bit of a story with Mongo DB.
So I, I first, yeah, I, I first encountered it during my
bachelor's when I became interested in, well, actually
MapReduce algorithms. So I wanted to do, yeah,
something back in the day, simple with Twitter data.
So like let's say, finding popular, popular hashtags or
(18:15):
similar, like analytics and Twitter data, it includes really
diverse contact content types, right, that evolve over time.
And then I've, yeah, stumbled upon Mondo DB, which was
essentially allowing storing this document based structure,
right, like various data types, rigid schema constraints.
(18:35):
And yeah, This is why, yeah, I'mjust started starting to work
with that. And interestingly enough, I
think it's also a good lesson for for me.
Perhaps it is like why? Why you should read
documentation because well, somewhere along the lines of
Mongo DB docs, it was 2015, I actually think.
(18:57):
So they were, it was written that guys, if you, if you're
trying to use this MapReduce forthis particular problem, you are
doing it wrong. Like there is like some
aggregation functions that you can make use of.
So like just just don't, don't use my MapReduce for that.
So, yeah, so I've spent quite some time like which which were
gone in vain. Well, and at least I, I learned
Mongo DB also learned a bit of Python because the core calls in
(19:19):
Python. So, OK, it was good.
And then actually just eight years later, yeah, 80 years
later, I think so when we started to work in our company,
we realized that our engine which was working fine in
academic settings, but when you bring it to the industry, right,
it requires many additional elements, right?
(19:40):
Servers, backlogs, data safety measures.
So, and then, yeah, I was just like, OK, what, what, what,
what's now out there. And then I came across Mongo DBS
again, as a, as a, as a, an interesting solution, perhaps
because you can imagine like for, for drones is 1 type of
data. But when you are like doing
finance or something else, it's a different type of data.
(20:03):
So again, like rigid schema. I pitched this idea to our CTO
Admitry and yeah, here we are. So basically it's flexibility
and performance characteristics made it well ideal fit for our
dynamic data needs. OK, excellent.
So the yeah, the, the I can understand why, yeah, in the
(20:25):
with the Lego blocks that you'reoffering and the different areas
in which you can work a rigid, yeah, rigid structure, rigid
schema wouldn't work at all. And, and to me, it's, it's
interesting to see that pathway,you know, that kind of
eight-year gap is because obviously in developer relations
where I work, we're, we're, we're about enablement and
awareness and planting the seeds.
(20:46):
But as I keep telling my VPS here, it's a long time
potentially for that seed to grow.
And in this instance you had familiarity with it and and when
the calling came, you went, I think Margot Abe could be very
suitable for this. What in particular, I mean, are
you taking use of time series data?
Are you using a Mongo DB Atlas vector search and and what we
(21:09):
have there now what what parts of of our ecosystem that you are
using at the moment then Albert?Right, right.
So I will, I will go into the details of how we use it.
At the moment we are indeed working with.
So we, we've started to work with this leverage in time
series database because it's a critical dimension in our
(21:29):
system, right? So basically Mongo DB allows to
well improve our ability to process temporal data
efficiently. And this is just, it's not just
technical optimization. I think it's fundamental to how
our agents understand and respond to the world.
And I must say that we, we haven't utilized this
(21:51):
functionality fully yet. So it's still something we are
working on and like perhaps actually we will jump on it
pretty quick pretty soon. As for vector search, yes, so,
and this is this is something with which is really cool in in
Mongo DB. So one of the pilot projects
(22:14):
that we've been working on it actually well, it it's we had to
work with a lot of textual data.So it's actually textual data
spread in time also. So it's like you can imagine
there's like some text that yeah, we have at well it one
hour time steps. So and then of course, well, our
(22:38):
agent well to enable this sort of navigating through these
documents, well, we did have to provide it to some, yeah, some
vector searches like what traditionally done in The Reg,
right. And this type of approaches.
So it's basically a way for. So vector search was a way to
(23:00):
navigate this embedded vector oflike that is embedding
representation of of the textualfiles which was working pretty
smooth. So unfortunately I can't show
that demo because it's under NDAbut but I will show another one.
OK, perfect. And you mentioned some customers
and some types of solutions, youknow the robotics, the drones
(23:22):
etcetera. Are there other industries that
you know can benefit from lazy dynamics solution as well too?
Is it, is it super adaptable, super scalable for other areas
or who are who would you consider your primary industries
that you would go after in termsof business within Lazy
dynamics? Right, right.
So I, I, I think that the ideal customer is like the
(23:46):
organisations that are building agents or autonomous systems for
as I said, mission critical tasks so that face significant
uncertainty in complete data andneed to make plans despite the
noise. So, and in fact, like currently
we are on boarding Co development partners across, I
(24:07):
would say 4 primary sectors. And like we are basically in
this stage where we're just trying to figure out, OK, what's
what's the most lucrative at themoment.
But it's so one is autonomous systems, yeah, particularly like
drones or wheeled robotics, thenlogistics, which is kind of
interconnected with that. And yeah, predictive maintenance
(24:29):
is basically to figure out, OK, when your reporting system or
when your, your system will failand like what, what is the
uncertainty around that that failure?
And yeah, there were quite some interest from finance industry,
but yeah, we are, yeah, not well, we'll see.
So and, and but but this interesting this as I said,
they, they share the common need.
(24:50):
So they require systems that canadapt to changing conditions
while maintaining this reliability.
Yeah. So that's what connects them
all. Which is, you know, it's, it's
nice to have that super broad audience.
And I think going back to the earlier conversation about
agentic systems, I think it's really early days to figure out
(25:12):
how and where they might be applied as well too.
I think we've all seen the clever demos, you know, but you
know, yeah, agentic system booking my holiday for me within
my budget in the and getting my travel itinerary.
Yeah, nice demo, right. But I'm really looking in all of
these kind of demos under the hood at, you know, where's the
(25:34):
production application, where's the scalable application, you
know, outside of these, you know, show and tell demos.
And I think you guys are obviously addressing that area
very much. So you seem to be at the at the
cutting edge looking ahead beyond this maybe what's the
longer term vision for laser dynamics at this time, Albert or
(25:55):
you know, kind of what else, what other areas could you get
into perhaps too? I'm conscious that we'll, we'll
show some demos. So I'm trying to just tee up the
understanding of how this could be applied perhaps?
Sure. I I, I think that just just to
linger a bit on, on the example you've provided with this, you
know, finding the, the trade, like within the budget.
(26:16):
I think it's, it's a great application actually of well,
it's a genetic system based on, you know, lambs, right?
Or like VLA, essentially Transformers.
But I wonder actually, are thereresidents where people actually
just drafted with their like a whole heart like to any of those
(26:38):
right. Like, yeah, I'm, I'm not.
So I I've seen quite some peopleactually well on boarding this
SDA SDRS, right, the sales development robots and yeah,
like a month my. Sales development robots as SDRS
like in Mongo DB, internally we have SDRS, but they're sales
(26:58):
development representatives. We haven't replaced them with
robots yet, have you? So you're seeing this?
Well, I, I do know that there's some people actually, they've
tried and they've, they've hatedit as well, right?
So, and the reason is the following again, right.
So you're basically, it's, it's not a compiloting system.
It's like a really autopilot where you basically bet your
(27:21):
business, you bet your reputation on, on, on the LLM
that can hallucinate. And and I said right, like for
this, for the spam, I can be perfect.
I would like for creativity, yes, but would you bet your
well, let's say life your business on that?
And I'm not so sure so and I don't know how our approach can
(27:42):
address that yet. I, I, I really like I'm we're
really in the scope of more likephysical world, like on, on,
yes, something like really tangible, some signals that we,
we, we know what is the meaning of them.
So and well, I'm not saying thatthis identic system that are all
now an owned well, what they're,which are now, which everyone is
(28:04):
talking right now about. It's, I think it's, it's, it's
it's a great deck for sure, but it's a bit explainable kind of
like it's a bit like you can't really figure out and trace back
OK, what what really happened there like, or if you if you can
do it, it can be really costly. Like with our system, it's well
usually like very much explainable.
(28:26):
So we actually do know all the time what went wrong in the
system. But because it's model based
approach, right? So it's, it's also does it I
know also to sell it is like this is the silver bullet for
identic system. No, it, it solves particular
problems within identic systems,But well, we'll see.
It's it's still a long way. And when you ask like about the
(28:47):
vision, I think the stretch goalis for sure to empower this
planning abilities of hygienic systems.
So essentially to become a brainfor well, autonomous systems or
let's say robotic applications. So I really want to envision the
world where autonomous system, they can navigate complex in
certain environments and yeah, enable this that the ability
(29:13):
decision making capabilities, yeah, that were previously been
impossible. So we can actually trust them,
yeah. OK.
OK. Well, I'm sure with that team
that you've collected around youin Laser Dynamics, I think very
much look forward to how things shape out.
Adita, thank you very much for the comments as well.
Appreciate that if there's any other comments, our questions
(29:36):
actually for Albert as we go through, please do let us know
in the comments and post them aswell too.
So I think like always, I try toget the guests to showcase
something and show us a little bit of a demo.
So I think given the high level conversation that we've had now,
Albert, I think it's probably about time to to showcase some
(29:56):
examples and a practical exampleand showing us kind of what
we've been talking about for thelast 30 minutes or so.
Right, right. OK, so let's start about that.
I'll just make sure that everything runs.
Yep, I think she'd be good. OK, so screen.
(30:17):
Yes, screen. OK, do you see my screen now?
Yes, that's it. It's, it's coming through clear.
We see lots of lots of super interesting stuff.
Albert here. What is it?
What are we? Looking at yeah, OK, OK, so so
this is actually a digital simulation of of the drone
environment. So like think of drone as an
(30:38):
agent and there are like 3 different tasks and scenarios
for for this drone. I'll start with what it's doing
and then how we do it actually. So so the first task as you see
here, so we have a drone which is basically continuously tasked
with going from 1:00 point in the well in this 3 dimensional
(31:04):
space to another one. The thing about this drone is
that it doesn't know anything about the obstacles that it will
see in front of it. So the idea is that OK, so the
drone is tasked. So you say, OK, drone, you you
should go there and then as longas it starts moving, it
encounters different objects andit needs to adapt in real time
(31:27):
to to avoid those. So as you can imagine, drone
doesn't have a luxury to stop. It's to pause.
It's thinking, you know, like, what would I do next?
So I mean, doesn't work like that.
So you really need to keep reacting on whatever
environmental changes occur. So here we're just basically
(31:47):
showing how drone sees its distances, like towards the
Yeah. So another yeah, the kind of
small example is that the drone here is essentially was not, not
it knows that it needs to to do this sort of simple pattern, but
what it doesn't know that it's it's raining outside.
So obviously when it's raining, this environmental condition
(32:10):
changes for the drone, so it needs to adapt.
OK. And that's what we what we show
here. So OK, there is sort of an
adaptation going on like as, as it's moving, so it tries to
learn also OK that it now needs to operate with its actions or
controls differently. So what does it have to do all
(32:31):
with Mongo DB? So and perhaps before going
there, so there is actually another front end.
So it's I mean, it's kind of real time demo of right the
dumping from so the the point where it needs to go generates
randomly and like as soon as it pops up, the drone kind of
figures out this action space where how it needs to control
(32:52):
its rotors to actually reach there.
So in that you can see it's all kind of happens real time.
So we can have the statistics over the engine forces.
There's also a bit of code how how it works.
But let me get first into the yeah, what what Ericsson, what
the Mongo DB has anything to do with that?
(33:13):
So so this is the demo that I just showed right where the
drone is just to call this goal.And there are three components
here. So the third very first
component is this front end is like basically your simulation
or in the real world, it would be just a real draw.
I reconfirm it's a, it's a server here which is running our
(33:38):
engine. Like you can think of it as a,
as a brain kind of that decides what actions sent to the real
world, the right. So it's kind of the inference
engine as we call it. And that's the open source
project that you gave me the link to earlier as well too,
This one here, OK. Yes, yes, yes.
So this open source, I can play,I can talk a bit on what's not
(34:00):
open source, but later on so, and then there is Mongo DB.
So why all matters? Because, well, all these things
need to communicate to each other and they need to
communicate each other pretty fast, right?
Because, well, there is of course the server and the front
end or like the real world wherethe communication needs to
happen real time. And we would also prefer to have
(34:22):
Mongo DB communicating to the server also as real time as
possible. So you don't want to have this
slacks and of course, and that'sessentially what's going on.
So there are like 2 communication pipelines.
So first is like from the serverto Mongo DB and what it logs is
current states data logs, right?What was happening with thinking
(34:44):
process, right? So how, how did we end up to use
that particular sequence of actions in order to get to this
place? We also store previous states,
the parameters of the drones, right, Because it could happen
that we don't have the full specification of of the drone.
For instance, we don't know it'sinertia.
So it's something we need to learn and we want to track over
(35:06):
time how the drone was learning that.
So this is very important and also for basically analysing it.
Well, if all of a sudden this drone flew to someone's else
window, right, we want to know what what actually went wrong
there and here here. What's important also is that
sometimes why this rigid schema comes into play, sometimes
(35:29):
sensors fail right? So it's something that the real
world can surprise you with, right?
So this can be like a a failure in the sensor in the road or
anything. And then it creates some null
values or like missing values. And that that's something we
also want to to log right away. And another pipeline is is from
(35:51):
server to the front end. So yeah, it's I see here it's
yeah, 3030 events per seconds per second.
And it's multi dimensional observation, right?
So we actually observe the droneobserves different well,
different data. So well, you can imagine like
it's from camera, it can be one type of data, right?
(36:14):
There is also, well, let's say the position from IMU sensors.
So it's really different than well, kind of scars data we are
working with. And yeah, so here what's going
on is like real time state estimating, prediction and
quality generation with Mongo DBas a back end database and it
(36:35):
works pretty well so. Good, good.
I'm glad to hear. Yeah.
And then yeah, just three things, what we are using Mongo
DB for as well. So we've talked about different
applications, some of which I'm able to show, but so for real
time embedded systems, mod DB turned out to be really good,
(36:57):
like for a thorough pool drive, minimum latency, fast date
updates. We also, as I said, like we're
working with unstructured data or like badly structured data
when you're working with texts usually like that.
So, yeah, there was a good use of Mongo DB vector search that's
(37:21):
also sort of a way for our agentto navigate the the environment,
OK to find, OK, what are the documents are relevant and well
predictive maintenance again, very similar, very similar to
real time embedded system. So we want to identify these
historical partners and react assoon as possible.
(37:42):
So yeah, as just to conclude, right, different patterns and
same database. So that's that's what we like
nonvidible for nonvidible before.
Excellent. I couldn't have said it better
myself. So no, I I'm glad.
I'm glad that you went through that.
That's that's perfect. We did get a question in from
from Janani and I hope I pronounced that correct.
(38:03):
And she's asking what kinds of path planning algorithms or
approaches do you use when the agents are involved and when the
obstacles in the environment areunknown?
Is that on yourselves and lazy dynamics or is that on the
client side when they use lazy dynamics?
No, that's a great question. Thanks a lot, Gianni.
(38:25):
So The thing is like, I would have to go into a bit of the
details of Eric's and Fer. So it's on our side, first of
all. So that's something we figure
out for for the for the client or like for the user,
essentially. So the way work and I really
advise John, I need to go into the open source or like reach us
(38:48):
out on the on meet hub. We can elaborate on that.
So first of all, we use we we caused this planning problem as
sorry for for for the slang, butbut as as a Bayesian entrance
problem. So we basically say as a the
planning problem can be reframedas a probabilistic entrance.
(39:13):
So think of it this way, right? So when you do planning, you
will basically say to your agentor the drone that you want to be
seen where in some time now you can like you, you can think it's
kind of inverse problem, right? So you're trying to figure out,
OK, what is the optimal, optimalpath?
However, why you can't solve it,for instance, with a star, which
(39:37):
is like a very common algorithm for path planning is, is that
the environment which we're in is dynamic.
There are many uncertainties around your sensors.
There are many uncertainties about other obstacles.
So we basically say to the agentthat it needs to dream about
this path 1st and usually when the when the agent dream about
(39:57):
this path and like that, that's kind of what we do in the
server. Of course it, it has very
optimistic agenda, right? So it thinks that well, there
will be no obstacles or maybe there will be yeah, well, there
will be no wind or something. So it tries to draw this very
simplistic path. Well, which goes ahead with this
Oh, with accordance to least acting principle, I will not go
(40:21):
into there, but sorry, sorry to mention.
And then? When the drone starts to move
according to its planned agenda,which we provide, which we, we,
we gave you this plan which you can execute, what starts to
happen is that well, there are like, well, some obstacles
appear or some failure secure and then it immediately tries to
(40:41):
well circumvented by replanning.And this replanning is usually,
well, it's it's a very expensivetask to do.
So there is a whole way of doingit in our way.
So we're we kind of developed this a smart way of planning,
let's put it this way. So we actually managed to replan
in real time. As for the yeah, As for the
(41:03):
obstacles, of course the the, the drone needs to sense
something. So there should be some sort of
a sensor that can say that the obstacles otherwise, well, we,
we are in trouble, right? It's like, yes, so, so there
needs to be a camera or like some, some lighter or whatever
(41:26):
that kind of can can tell you OK, how far you are from from
the office. OK.
And forgive me if you've alreadytalked on this, but the that re
planning obviously that's all automatic.
Is there some feedback loop backto, you know, back to the
customer who's using this that that has occurred?
Or is it go and do the job, takeinto account of the paths, the
(41:48):
obstacles, the environmental conditions and finish the job.
And yeah, so, OK, so this is where open source differs from
progression. So in in open source version,
you got to figure it out yourself how you're going to
utilize all this replanning and planning and executing on the
plan together. And I mean, we're, we're helping
(42:13):
with that, right, as, as much aswe can.
Like on our, well, we didn't have examples or GitHub
discussions, but it, it's, we also recognize that it's hard
and that's actually where the value position of laser.
And so we automate that. So we basically tell you, OK,
forget about all these worries, how to sort of create these
(42:35):
loops or feedback loop as you just mentioned.
And we will just do that for you.
So all you would have to do is to pre specify essentially to
call just three functions, whichis plan, act and and learn.
So and and all the rest. We are we are doing this magic
for you. So yeah.
OK. Oh, brilliant.
OK. So that's the pathway from open
(42:57):
source to to revenue as far as lazy dynamics are concerned then
so. Right.
OK, we'll look it makes a ton ofsense.
It's a well worn pathway for forservices.
Mongo DB is open source essentially, but if you want all
of the the enterprise features or if you want to use Atlas in
the cloud, it's it's our software as a service platform.
So that makes sense as well too.So it's a it's a well worn
(43:21):
pathway in Ireland, where I'm based, we have a company called
Mana Aero that does drone deliveries and you know, it's it
seems amazing. It seems fabulous what they do.
You know, it's, it's for low value items, you know, copies
and fast food and takeaways and things like that too.
But I, I can imagine, given whatyou've discussed, the
(43:41):
complexities that they must haveto overcome the particularly the
environmental stuff. As I said to you when we were
jumped on the call, it's, it's areally bad weather day here in
Ireland. I can imagine that that affects
the flights and, and paths of those type of services.
Are you seeing those services, they always seem to be kind of
trial. Are you seeing those
(44:02):
commercially used? I know there's been some nice
case studies in in, you know, Africa and elsewhere for medical
supplies deliveries as well too.Is that any of your customers
today? So As for the robotics right
concerned we're we've just started to onboard partners.
(44:22):
So I can't really disclose them.So we are, but they're like fast
like forward thinking customers who who recognize that and then
we we, we are providing this sort of a solver for their drone
dynamics. So and I, I do, I do believe
that in the future, that's how the yeah, mass delivery will
(44:47):
operate. But there are quite, quite many
issues that needs to be addressed right.
So well, a lot of liability issues with that then there is
obviously this well sound noise,right, this this pollution,
right, this pollution. So this is something that yeah,
(45:10):
I think this drone deliveries will will face.
I know there are companies whichare actually trying to well
develop the drone such that theybasically are almost silent.
So but that's, I think also a few years away.
I, I remember when I was doing my masters in, in, in Japan, I
think Rakuten was also just about to launch their drone
(45:33):
delivery program in. I've been there again.
I've been there again, like 3 weeks ago, it didn't
materialise. So yeah, I think it's hard.
It's hard. And as you said, there's a lot
of uncertainty around those deliveries and like how also
people approach this drones, right?
So it's there, there are many issues that if you look at like
(45:57):
from kind of surface perspective, seems like wow,
it's, it's easy stuff like you figure out.
But yeah, the devil is in the details.
Yeah, very much I think, yeah. As you said, there's a lot of
these companies just about to launch commercially and and it's
always just about to launch. You mentioned I think at the
intro as well, probably the moreunderstandable scenario of, you
(46:20):
know, warehouse and there's robots and there's automated
systems and warehouse that do things.
We see from time to time the large retailers, you know,
automated warehouse, etcetera. Obviously the drone example is
one that catches the attention, but I would imagine the scale of
the warehousing robots automation is something that's
an enormous challenge to tackle as well too.
(46:40):
Oh yeah, yeah, I. Thought Lazy Dynamics can offer.
Yeah, I think, I think I can also sure share that.
I think we'd have time, right? So.
Yeah. That's.
OK. Janani came back and she said,
you know, she learned some probabilistic path planning in
class and why she asked the question.
So she she thanks you for your insights there, Albert.
(47:03):
Thank you. OK for the question.
Yeah. Thanks a lot.
Yeah, yeah, a lot of these questions for from people who
already know the answer, right. So, but yeah, OK, good, good.
Now I have yeah, I just invite her to to check us out.
So I think she would have it. Interesting.
Yeah. So, yeah, there's a small, just
another example. I think you just mentioned this
(47:24):
warehouse. And I think this is also perfect
kind of situation where we, our solution is very handy.
So here you have this kind of, yeah, like a bird's eye view of
a warehouse. So you have these four agents
that are tasked to go from one location to another and
(47:48):
locations are represented as stars.
This small agent, they also havesort of this, you know, this
death circle around that, which is basically their well, their
sensor like where they can locate the obstacle and the task
by itself. If it was just a static
environment, it's not that hard.Like it's really not that hard.
(48:11):
I mean, especially well, it's it's static environment only
robots are there. But imagine it's not only agents
like I mean robots, but they're also humans.
And and not not to say that onlyhumans make mistakes like also a
genetic systems like robots alsocan well block the aisle or do
something nesting like that willdisrupt the the plant.
(48:34):
But here we actually do is a very, let's say hostile
environment, like basically an environment that changes all the
time. So when the agent tries to
execute its idealistic plan, right, like is it is it tries to
sort of find to to dream about how it would go that the the
(48:55):
environment always pushes back. It doesn't let the agents to
materialize. So they actually need to
navigate seamlessly. OK, how do I avoid this?
How do I avoid that as they go? Because yeah, again, we we can't
stop some of the processes whichare ongoing.
So the task was here to, OK, we define the requirements for the
agent. That's actually example.
(49:16):
I think also now we're open source, which is I think like 15
lines of fault just to define this model.
And the the task is the following.
So respect the speed, respect the distances and go from one
location to another. And then, yeah, the agent, they
figure out how to do that. OK, OK.
Yeah, an incredibly, I suppose under the hood, we don't
(49:37):
generally think of these things to this level of detail.
But obviously, of course, and, and as you mentioned the the
driving example originally, you know the environment is ever
changing and and you just can't,as you say, I love the term that
you use dream about the perfect path, but that never probably
ever occurs, right? Yeah, absolutely.
(49:59):
I mean like, like, like for us humans, right?
I mean, just saying you were like, do you think that
something will occur in like andit just happens in a completely
different way? And nevertheless, yeah, we
managed to adapt. So I mean, I, I haven't
mentioned that, but like actually our professor, right,
like he was, he's been always fascinated about how biological
systems adapt. And like many of things that
(50:22):
we've implemented also actually tried to draw some of the
inspiration from what he's been thinking about.
So yeah, how how we human adapt reactively as the environmental
environment changes. So that's that's an interesting
topic for sure. Definitely, definitely.
Oh, this has been fascinating, Isuppose.
(50:42):
And look, and obviously that thefact that you've done a
doctorate etcetera as well too. You're an avid learner, Albert,
but you know, this space as you,you know, look for those that
who don't know anything about this space, you think AI is only
2 1/2 years old, right? For anyone who does, it's been
machine learning and artificial intelligence for, you know,
many, many decades. How do you keep up to date with
(51:05):
the changes in this space and and where do you go to to learn
and keep ahead of things for Lazy Dynamics as the CEO?
Yeah, that's a great question, Shane.
I I should say that I think thatwhat, what's really important is
to build the community around you that led to different things
(51:26):
specifically. Yeah, I have friends who are
thinking that, yeah, Transformers is this best thing
since sliced bread rice. We're really like very like
venturing into that and I respect that I, I think that
even though, yeah, I, I disagreewith them, I'm trying to keep up
with that. So I, I think that for us,
(51:46):
what's important is to really build up the community around
things which are complementary or competitive to us.
So just to stay, to stay kind ofrelevant to see what's going on.
I think there's also a lot to learn from these things, right?
So you can, you can have all this other agentic systems,
(52:07):
right? There's a lot to learn there.
So from realization perspectivesand so forth.
And so this is basically the waywe, we stay relevant.
And I think what's most important is for, for our
companies that we, we are reallyventuring in.
So what we do, it's very important to believe when what
you are doing right. So it's of course not this
(52:29):
pretty shared kind of stuff, butthat, that that's the reality.
So we, we do think that our approach is, is the best one, is
the right one. And like for specific problems,
it will be, well, it, it will draw others obsolete.
And I think that this is very, this is very important.
So yeah, I guess these two things, community and for
(52:52):
believing in what you're doing, yeah.
Yeah, I love the community aspect of that.
And and I suppose obviously you're, you're getting feedback
from the early customers, the early adopters of Lazy Dynamics
as well too. And you're enabling them and and
growing with them as well, whichis, which is amazing to see.
It is a yeah, it's, it's a fascinating area that you've
(53:13):
you've gone into. And I think obviously timing is
everything too, Albert, Right. So you were, you were doing work
in these spaces, you know, throughout for a vast number of
years obviously and and then sawthe commercial prospect of what
you've been doing and gathered that team around you.
Yeah, that's right. That's right.
(53:33):
I think, yeah, I shouldn't have mentioned the partners and
customers because this area the well, the real asset who, who
kind of makes this reality checkfor the product, right?
So because again, like as, as founders from academia and, and
they see it a lot, right? Like everyone who is trying to
launch out of the lab, they're trying to basically they have an
(53:56):
amazing technology and they're like trying to find the fit for
it. And, and, and we were like, just
like that. So with that, yeah, we have this
amazing technology, but I mean, truth to be told, no one cares,
right? So it's about the the right
problem that your technology cansolve.
So, and this is hard and like assoon as you realize that you
(54:17):
find your customers and you actually get very valuable,
valuable feedback actually that shapes, shapes your products,
shapes your technology in the right way.
Yeah, no, it's great. And I suppose given the
background that you all have with all your PhDs behind you
and all the different areas, I, I used to work with a lot of
(54:39):
startups And what we find a lot of the time as startups are
building products, but you know,they're, they're afraid to go
out there into the open and get customers because it's, you
know, it's, it's a real world validation of their system.
Whereas you seem to be very clear on, you know how Lazy
Dynamic can help these customersalready and the environments and
(54:59):
the applications to which they can be beneficial, right?
Right, right. Yeah, That's that's, that's a
great conclusion. Yeah, I think that's.
Yeah, well, like for for a company 18 months or so old, I
think you've come on in in leapsand bounds Albert.
And it's, it's great to Share your story and the the journey
that you've had and give us kindof the glimpse of the insights
(55:22):
into kind of the innovative worlds that you occupy in Lazy
dynamics before we wrap up. And and I didn't see any other
questions come in. So obviously, as I said to you
on our prep, usually that means you've been very clear in your
mission and what you set out to explain here.
So that's that's great. Any last words for our audience,
maybe from the point of view of,you know, going out there and
(55:45):
found in a company or, you know,what's happening in the identic
world or how to keep up or how to learn?
Yeah, I think we've touched uponall of these topics.
I think that, well, there's justone thing I would, I would love
to to ask people for, I guess isjust to check out Oryxin Fer,
(56:05):
check out Laser Dynamics. Oryxin Fer is completely open
source and I'm, I'm sure that everyone, every expert in their
own field would be able to relate to some of the problems
we highlight on our open source website.
And well, just reach out for, for, yeah, for anything like it
(56:28):
just be the open source questionor like the possibility for
potential project together. So we're open for any sort of
collaboration. Yeah, that's that's basically
it. Excellent.
Well, listen, I think that's a good call out to for to wrap up
our our podcast here, Albert, it's been a pleasure having you
(56:50):
on board and sharing. I will certainly keep an eye on
what lazy dynamics are doing. And you know, whenever the next
stage comes around or whenever the next demos that you have,
I'd love to get you back on the show as well to to explain.
But I think for now, obviously, as you said, go to or X infer
and also as you see on the screen there at the moment, Lazy
(57:12):
dynamics.com to learn more. But for now, for me, Shane
McAllister, thank you so much, Albert.
It's been a pleasure to have youon board and all the best with
you and the team at Lazy Dynamics for the future.
We wish you every success. Yeah.
Thanks, Shane. Thanks.
Excellent. And for all our viewers, thank
you very much for joining us anddo keep up.
(57:34):
As I said in the introductions, if you like and subscribe to the
Mongo DB YouTube channel, you'llget alerts for similar live
streams such as this. And also follow us on LinkedIn
to get similar event notifications there too.
I stream most Tuesdays at roughly this time with lots of
amazing guests such as Albert. So do join in.
But from both of us now, thank you all for tuning in.
(57:56):
It's been a pleasure to have youand Albert.
Thank you once again, it's been great.
Take care everyone, good luck. Bye bye.
Bye.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com
The Bobby Bones Show
Listen to 'The Bobby Bones Show' by downloading the daily full replay.