Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:00):
Welcome to Tectastic, where we navigate theintersection of technology and business,
uncovering innovations that redefine our world.
Brian Du Boyd joins us today on it'stechtastic, and he's got a bong background in
industrial AI.
And this is a space that I personally findvery, very interesting.
(00:22):
So I wanna say, Brian DeBoyd, welcome to it.
It's fantastic.
It is lovely to have you here.
Thanks, Christian.
Yeah.
Awesome.
We'll be here.
So there's a couple topics that I think thatare very interesting and related.
1, industrial AI and advanced manufacturing, Iwould assume, means, like, 3 d printing CNC
robotic assembly that hit the thing.
You're probably an expert there.
Yes.
Somewhat.
(00:42):
And it's interesting because, you know, to bean expert in all those things you just listed
is tough to find one person.
There's definitely ask fix that of that that wedo.
We do some CNC.
We do some robotics.
Actually, in a lot of ways, that industrial AIthat that we're gonna be talking about, is
actually applied in, the more traditionalautomation spaces of of the manufacturing
(01:03):
facility.
So I don't know how familiar you are, but youknow, if you think of PLCs and some of these,
more traditional automation that's been aroundfor a long time, we can actually apply that
individually on top of those existing
Oh, no.
So there's a big need there.
Yeah.
CNC is very specific.
It's actually kind
of
a specific niche within the broadermanufacturing world.
(01:24):
We do, like I said, we do some of that, butreally if you focus more on manufacturing lines
as a whole.
So if you've ever seen the show how it's made,that's typically what I refer people to.
Right?
So I watched that growing up.
Loved it.
Really knew nothing about the manufacturingspace before I moved into this 24 years ago,
started my current company, Robisys, 24 yearsago, and learned everything, you know, about
(01:47):
this manufacturing space.
So while robotics certainly plays a a role incertain industries, In other industries, it's
piping, it's big reactors.
If you're making chemicals, if you're making,life science ingredients, pharmaceutical
ingredients, if you're making, food andbeverage type products, there's less robotics
there and there's more just conveyance andyou've got big, you know, again, reactors and
(02:09):
vats of things and piping, you know, materialand things like that.
We can actually apply industrial AI to solvethat.
And by the way, the same labor shortages thatthey're seeing on the CNC side that you
referred to, we are absolutely seeing thatacross the board in manufacturing.
So I had a particular passion for advancedmanufacturing sharing.
I actually have a patent out there for a longtime ago in the space.
(02:31):
I haven't I don't think I've ever shared thison the podcast, but the the year between high
school and university I went and worked for ahigh-tech fab on the graveyard shift in their
prototyping lab for what would become the theface change, printers for Xerox by the time it
was Tektronix.
Okay.
(02:52):
And that gave me a whole bunch of insight intohow, like, that type of printing was done, and
I started applying that to 3 d printing and allthat type of stuff.
You're not exclusively talking about 3 dprinting, but I have a fascination with it
because a different problem that we've tried tosolve repeatedly can be addressed with it.
And that problem is this, and I I don't knowthat we'll spend a whole lot of time talking
(03:12):
about it.
But as the world moved from solely informationbeing online and basically all physical goods
being available online.
There was a transition based that that wentthrough, but there was a forcing issue that was
becoming a reality for anybody in direct toconsumer.
And that was, I want it now.
Yeah.
Right?
Yeah.
But I think it's gone beyond that.
(03:32):
I think I want custom and now You can solve theI want it now with logistics, with a very
intelligent AI driven logistics, but you can'tsolve the custom problem.
Right.
It sounds like that was probably quite a fewyears ago.
It was.
Yeah.
Manufacturing has advanced significantly sincethen.
And one of the and I'm not an additivemanufacturing expert, but I do monitor these
(03:52):
trends.
And and one of the areas in particular thatwe're seeing huge improvements there is not
even the direct to consumer route.
But think being able to 3 d print majorcomponents of airplane wings.
And so in the past, you only had limitedoptions.
Right?
You could see and see these parts you couldcast them or you could forge them.
I mean, those were basically your options, andall of them have pros and cons, particularly in
(04:16):
really complex parts.
And so what ends up happening, it is you end uphaving to break a a large complex part into
several smaller parts.
Each of those then has different failure pointsbetween them and the connecting all of that
back together.
With it, with additive manufacturing, they canactually print that all as one single piece.
And so there's there's a lot of research goinginto that.
(04:36):
We actually, worked with a company that wasexperimenting with making the, the metal powder
that would eventually be used in these big, bigadditive manufacturing facilities.
And by the way, metal metal powder is reallydangerous because it can explode.
Yeah.
And so that was a whole that was a veryinteresting project to work on.
But I yeah.
That's not my area of expertise, but I thinkthere's a lot of really interesting things that
(05:00):
that are gonna be coming down the pipe.
For additive manufacturing.
The the additive manufacturing part, like Isaid, is a bit of a passion, space, and I spent
a lot of time screwing around with it.
But it's a hobby more than as a, like, career.
Right?
I went hard core down the software side of theworld into AI.
So the industrial manufacturer sharing clientAI to, I'm assuming processes to help do
(05:23):
improvements and efficiency gains and all thatkind of fun I really wish that one of my former
colleagues who, worked with a couple differentcompanies was scared, Diana, She did process
improvement for some very large companiesaround their manufacturing process because
that's where she's a process engineer.
And she would be the best guest like, to topair you up with on this to, like, ping back
(05:44):
and forth.
Yeah.
I wanna hear more about it.
I wanna understand it better because, like,when we were talking at Wayfair around a couple
different problems.
One of them was the custom and now and infurniture and home goods.
That's like a really hard problem to solvebecause It's hard enough to get enough sofas in
the door to be able to have a big catalog.
It's a whole another thing to be able to have aadvanced manufacturing facility that can do all
(06:07):
the pop simple things that are involved infurniture.
But to even get to that step, you first startrefining the process.
And with, like, Nike, with the ID product, thecustom shoe you could design online.
That was fully figuring out the process acrossmany different facilities to try to do
backwards.
What can we offer the consumer?
And when they start making selections, how dowe limit them to only the things that that
(06:30):
facility can manage?
Right.
You know, when you talk about processimprovement, this is very much in in lean and 6
Sigma in these existing process improvementmethodologies that have been around for a long
time.
Industrial AI is is another tool in thattoolbox.
So we, you know, we try to say to ourcustomers, like, this isn't, you know, the
silver bullet.
It's not gonna solve all your problems by anymeans.
However, What I've seen is is that we can applysome of these more advanced industrial AI
(06:55):
techniques, which I can talk about here in aminute.
We can apply them two problems that havetypically been unsolvable in the past.
And what I mean by unsolvable is is that youtypically would just have to throw manpower at
a particular problem.
Or it was just a problem that everyone justkinda lived with, and it was kinda the state of
the art of that manufacturing, you know, makingthat particular type of product.
(07:16):
And nobody really knew how to solve that.
You know what I mean?
So it was just that was that was just how youdid things.
And we're now able to pack those types ofthings.
And one of the advanced types of industrial AIthat that robotics actually specializes in, and
this is cutting edge stuff.
It's called autonomous ai.
So this is where we're actually able to trainAI models to make human like decision making on
(07:41):
the plant And that is where we've been able tosolve these unsolvable problems.
That's where we've been able from a processimprovement standpoint been able to squeeze
out, you know, single digit percentimprovements that equate to 1,000,000 of
dollars every year, online that maybe havealready been optimized and they've gone through
66.
And it's like there's nothing more you can doto this line until the a technology like this
(08:04):
comes long, and now suddenly we can makeimprovements that have never been able to be
made before.
One of the other interesting aspects of this isthat there's also kind of a flow through to
sustainability, right, because if we can reducescrap or we can reduce energy consumption, yes,
the company makes more money, but it also has adirect impact on landfills, and it has a direct
(08:26):
impact on wasting materials, and it has directimpact on wasting energy.
All of that then flows through from asustainability standpoint, which is a big win
because a lot of these companies now do havesustainability goals too.
I mean, when they look at trying to hit some ofthese aggressive sustainability goals, even
with all of the existing process improvementmethodologies that are out there, you're not
gonna get to 2% per, you know, improvement onenergy reduction and things like that.
(08:51):
Unless you look at an advanced technology likeindustrial AI.
So that autonomous AI that we're talking about,the underlying technology behind it is a
technology called deep reinforcement learning.
Now you're plugged into this world.
Is that something you're familiar with?
Yes.
I I am, but you should explain to the audiencethat you like.
For the for your audience.
So deep reinforcement learning came out, deepmind, which was a Google spin off.
(09:13):
And around 2016 ish, they had a program comeout called Alpha Go.
It beat our best Go Grand Masters they did thesame thing with alpha 0, which beat our best
chess grand masters.
Then they did the same thing with alpha starand it beat our best starcraft players.
So they were on to something.
And some very clever people said, well, okay.
We can take this technology that's kind ofright now, it's being used to play games, and
(09:37):
we can actually apply it to the industrialworld.
So now fast forward, you know, almost a decadelater, and we now have customers that are
leveraging this deep reinforcement learningtechnology to make human like, strategic
decisions, which is really that strategy aspectis one of the key things that sets DRL apart
from other AI approaches, be able to make thosehuman like strategic decisions on the plant
(10:01):
floor.
And that we're able to apply that, and I lovethis about it.
We're able to apply autonomous AI to solveproblems as varied as actually controlling the
control system.
So actually making the product and turning theknobs just like an operator on the product,
more on spec product, all the way to productionscheduling, which is a notoriously hard
(10:22):
problem.
And again, when you talk about trying to get tothis made to order, made to custom type of
world, specific and specialized customizationproduction scheduling becomes one of the
biggest bottlenecks to that, and we canactually apply this to solve those types of
problems as well.
One of the other things that just keeps poppinginto my head is the non formity, manufacturing.
(10:44):
Any two lines doing the same thing for the samecompany can do it fundamentally differently
because past equipment purchases or whatever.
So do you have to do domain reinforced learningmodel for line type situation, or is it like
you're deploying 1 and it kind of understandsthe differences.
So it just depends on how close the lines are,but in a lot of ways, you know, once we get a
(11:06):
working simulator, we can oftentimes just dosome quick modifications to the simulation get
it working for the second line.
And then we've got, you know, all of thepipeline in place and they're all the project
in place to be able to retrain that brain onthe second line.
So it depends.
Some of our customers are, at least withinlines within the same facility.
(11:28):
They've been really good over the years atkeeping all the equipment the same and sure,
you know, because they also have to maintainthese lines.
So there's they have had incentives over theyears to try to keep stuff as standardized
possible.
And obviously, they now reap the benefits ofthat if they were able to do that because now
we can train once and deploy at multipleplaces.
But certainly between facilities, we absolutelysee that non standardization that you're
(11:49):
referring to.
So it depends.
But we typically can get a good amount ofreuse.
There's another aspect though that really makesthis reusable that's worth emphasizing.
And that's that this autonomous AI approachisn't just about DRL.
That's why I don't just refer to it as DRL.
There's another really important aspect, andthat's what's called machine teaching.
And machine teaching is where we actually sitdown with the subject matter expert We extract
(12:13):
from them their expertise and we actually bakeit into the model.
So we actually build a workflow that representstheir heuristics.
Some of them, I mean, maybe not even be writtendown.
It's just like, oh, yeah.
When I see things when I see this RPM kindagoing this way and I see this that I know I'm
in this state, so I'm gonna do this versus if Isee this RPM going low and I see these other
(12:35):
things happening, I'm in this state.
I know I'm in this state.
I'm gonna do this.
We build a workflow of all those decisions.
And then we actually hand that to the DRL, andit actually trains around that.
It's still able to find all kinds of ofinsights that a human wouldn't have thought of,
but that actually guides its training and thenice thing about that is that's almost always
(12:56):
reusable across lines, regardless of theunderlying equipment and things like that.
I was gonna say that part alone is really,really valuable.
Oh, yeah.
Because you can have same role, same line, youknow, same function, different shift and that
they're doing things just a little bitdifferent because they haven't picked up that
specific piece of knowledge that they need.
(13:16):
Right.
So the variability that you're gonna havebecause they're humans involved in operating
this.
But I then I wonder, though, if the way thatyou get information out of people itself ends
up being a particular problem because there's Ican imagine a simulation that they're involved
in.
Right?
And as they're as they're they're responding tothe emulation.
They're they're basically training your model.
(13:39):
But a question and answer thing, like, you haveto know what questions to ask.
Oh, yeah.
Oh, yeah.
I mean, it helps that we've been at this for along time.
So we're Yeah.
We have a lot of expertise, deep expertise,here at Roevises across all the different
industries we work in.
And we work across 14 different industries.
So, if we're going into a life sciencecustomer, we know generally the process and we
(13:59):
know enough that we can sit down with an for anoperator.
And so we know generally we could write thatworkflow ourselves without probably having that
interview.
But what we're trying get out of the interviewis let's get real specific about this part of
the process and what are the heuristics andwhat are the tips and tricks that you use to
squeeze the absolute most out of this part ofof the process.
(14:20):
And that's really then what we wanna give askind of a starting point to the autonomous ai
so that while it's training, it at least knowswhat are the best practices.
As you pointed out though, that's a reallyinteresting aspect of all of this because that
is what a lot of our customers are interestedin because those experts are retiring.
So as as an example, I have a vinyl extrusioncustomer and I was visiting.
(14:42):
I was to the plant manager.
And he said, you know, I've I've got 2 guys whoare absolute experts are running this final
extrusion line, and he goes, both of them are 5years from retirement.
He said, I don't have a bench.
He said, I've got 1 and 2 year guys.
I've got high turnover in that position.
Nobody really wants to spend the next decade oftheir life learning how to run a vinyl
(15:03):
extrusion line.
And that's just kind of the way it is rightnow.
And we can lament that and we can wish that itwas different.
Or we can just embrace the fact that that's thesituation right now is that nobody wants to
spend the next 10 years learning how to runthese lines.
And so that expertise is in the heads of thosetwo guys he knows that as soon as they leave,
his options are either adopt AI or live withthe fact that the line's never gonna run as
(15:25):
well as it did, you know, with those experts.
So This gives, you know, those folks anopportunity to extract that expertise from
those SMEs and bake it into these AI systems.
You know, now this thing can can look over theshoulder of a 1st year operator and say, hey,
if you just turn this knob 2% more to theright, you're gonna squeeze this much more
(15:45):
operation out of the machine.
And what's interesting is in talking to thoseSMEs, you know, a lot of people would think
that they would feel threatened by that.
It's the exact opposite.
These folks have spent their life learning howto run this equipment.
This has been their passion project, theircareer, and this is what they see as their
legacy.
Right, is baking all of that expertise in.
(16:05):
It's funny.
I I gave this example once.
I said, you know, if I find one of those guyswho's an spirit at running a vinyl extrusion
line.
And I've spent the last 15 years of my lifelearning how to squeeze every little bit of
performance out of that piece of equipment.
You know, vinyl extrusion.
It's a large complex process, big, bigequipment.
And I know everything there is to know aboutthat.
Who am I gonna talk to about that?
Right?
The my buddies at the bar don't wanna hear aword about it, and I can I can guarantee you my
(16:30):
wife doesn't wanna hear anything about it whenI get home?
Yeah.
And yet I sit down You know, my me and my teamsit down with this this expert with humility.
And we say, tell us everything you know abouthow to squeeze every how do you do it?
How do you do it so well?
They light up.
Like, they're so excited to talk about this.
No one's ever cared before.
And so we're able to extract all that, bakethis into the AI and then be able to build a a
(16:52):
human like decision making, neural network thatcan then sit on the line and be able to advise
the next generation of operators on how bestto, to operate that line.
I can think of a couple things that would makeme, a little bit nervous about tech knowledge
like this.
My company does AI also, and we're solvingweirdly kind of a similar problem, but in a
(17:13):
totally different place, software development,And one of the things that concerns me about it
is that it it's like dumbing down of the peoplethat are doing the role and how we were taking
away the parts that we think of as innatelyhuman, and we're giving that sort of a shame.
And now the human is doing the parts that seemalmost like the machine part, like, turning it
(17:34):
off because a different range will be.
What worries do you have about what you'redoing, or do you have any concerns about the
future based on what we're coping.
Well, I'm kind of a natural warrior anyway, soI have lots of worries.
But, I would say that, yeah, I mean, So it'sit's it's a general worry that I have anyway
about anywhere from autonomous driving toautopilot in planes.
(17:57):
So the more capable that we make a plane ofbeing able to fly itself, the less muscle
memory we're building in pilots to be able tohandle those weird one off 1% of the time
emergencies that can happen.
And as we saw, there was that Air France, Ithink it was playing the crash in the mid
2000s.
And it was a fully functional plane.
(18:17):
Literally nothing went wrong.
And what happened was the transition autopilotto a human pilot.
And it was a less experienced pilot.
The experienced pilot, had actually, went tosleep and turned it over to a less experienced
pilot.
And it's like, just fly the plane.
Like, you just have to fly a plane.
Keep it straight.
And they had, one of their tubes froze and,very natural thing happens all the time.
(18:37):
The autopilot switched over to manual.
And that transition from auto to manual, it'sthe same thing in autonomous cars.
That transition from auto to manual is a veryconcerning area.
For me, especially moving forward.
I don't think there's anything we can do aboutit as a society because, again, like, that's
just where we're at.
We're going to have this I think a lot morefocus needs to be put there.
(18:58):
We spend a lot of time talking with operatorsabout, okay.
When this thing doesn't know what to do, it'sgoing to flip control back to you.
Are the scenarios where that can happen, here'swhat you need to do, but it's it's a challenge.
And that's something that we're gonna face.
And I think, unfortunately, we're going to seemore since like that Air France, crash where
everything's fine.
You just have to fly the plane.
(19:18):
But because of someone had
not spent enough time in designing that system.
There was errors flying everywhere that were itwas alert overload and there was panic that
happened with the pilot, and that's what causedthat crash.
And I think we're gonna see the same thinghappen with these autonomous ai cars.
Everything's going fine.
Just drive the car.
You know how to drive a car.
(19:38):
Just drive the car, but that transition There'sthere's errors going off, there's alerts going
off, and suddenly you're in control again, andthat jarring effect, I think, is is very scary.
So I'm I'm I've actually spent a lot of timethinking about that, and I'm very nervous about
that.
But the good news is is that in themanufacturing space specifically, we've got
lots of safeguards that we can put in place sothat that is really not as big of a concern as
(20:01):
we're gonna see in other areas.
But broadly speaking, that transition pointbetween automatic and manual is something
autonomy and and manual operation is somethingthat I I am concerned about broadly.
But again, I spend my whole life.
You know, my whole day is thinking about AI andthings that's, of course, something I think
about.
So
It it's a fascinating phenomena, by the way.
So if you talk to an engineer who designselevators for a living or designs the a
(20:25):
airplane themselves, would you use the productto be designed?
100% every engineer says yes.
Here, would you get into a plane that wasrunning the software that you rode to fly
itself?
Would you fly in that plane?
I'd say it's close to 100% would say,absolutely not.
Yes.
Yeah.
And I have a software background as well.
I have a computer science degree.
And, yes, like, for some reason, a softwarefolks just are we're always concerned about all
(20:50):
those little edge cases that that could pop upand yeah.
So The unknown unknowns of tariff.
Yes.
And the that's impossible.
To calculate for all of them.
And so you know, just in your head, you'relike, I didn't think of everything.
I'm not that smart.
Right.
Yeah.
Yeah.
I think that, I mean, we're getting realphilosophical now, but I think the difference
(21:10):
there is is that engineers have the benefit ofof physics.
And so they can have, multiple layers of kindof defense against those, Swiss cheese Pride
problems or whatever those, you know, thoseblack swan type of things.
They have multiple of defense.
And, software tends to have only single laserdefense.
And honestly, that's something that I think assoftware development matures as an industry, I
(21:31):
mean, we're nowhere near what engineering interms of number of years.
Engineering goes back 100 years.
We have, you know, you know, what 80 years,maybe, not even in software development, years.
So once we as we continue to mature as anindustry around software development, I think
that, we need to learn those lessons and startto build in all those multiple layers of
sophistication and in protection, and defensein-depth and things like that in the
(21:54):
cybersecurity world.
Those are the types of lessons that we need tocontinue to learn.
Brian, an absolute pleasure, man.
That was fantastic.
Thanks, Christian.
Appreciate it.
And that's a wrap for this episode ofTectastic.
Wanna thank you personally for joining us, andwe'll see you next time.
Until then, keep exploring and stay curious.
(22:17):
Thank you for listening.
If you are new here and enjoyed the content,please subscribe.
It really helps us out.
And if you are a regular listener, thanks somuch for your continued support.
Overwhelmed by tech debt.
Discover Vala AI, the solution to techchallenges with the simplicity of a click.
No engineering background?
No problem.
(22:37):
Vala AI enables anyone to effortlessly tackletech issues, freeing up your time from tech
headache.
Make tech debt vanish with Vala AI, where yourtech solutions are just a click away.