July 10, 2025 • 25 mins
This episode of Auto Collabs gets seriously smart as Paul, Kyle, and Michael welcome Sanjay Varnwal, founder of Spyne. With a background that includes time at Amazon and deep roots in AI-driven product development, Sanjay shares his journey from general e-commerce merchandising to building ultra-specialized solutions for the auto industry. His company’s obsession with perfection—like making sure side mirrors don’t vanish in digital photos—isn’t just nerdy, it’s the kind of tech precision that can radically improve dealer efficiency and customer experience.
Sanjay also breaks down the evolution of agentic AI—technology that doesn't just talk, but acts on behalf of dealers to optimize pricing, image performance, and even ad budgets. It’s the next level of automation, where your digital teammate gets smarter (and more autonomous) over time. Whether you're deep in the tech weeds or just trying to understand how AI can lighten the load in your dealership, this convo offers sharp insights and a clear look at what’s next.
Takeaways:
0:00 – 🎙️ Intro banter: Meta Ray-Bans, Weezer dreams & Wiggles hype
1:48 – 💡 Introducing Sanjay Varnwal from Spyne
2:28 – 🤖 Sanjay’s first AI experience at Amazon Go
3:39 – 🧠 How AI’s been around long before 2022
5:02 – 🎥 The evolution of AI in media and merchandising
7:40 – 🔄 How AI is changing human behavior and productivity
9:45 – 🚗 Why Spyne focused on auto merchandising
11:30 – 🛠️ From general e-commerce to auto-specific solutions
13:05 – 📸 Solving hard problems: lighting, stabilization & studio-quality images
15:46 – 🪞 The vanishing mirror problem and how Spyne fixed it
17:40 – 🧑💻 Building an AI-powered teammate for dealers
19:04 – 🧭 What is agentic AI and why it matters
20:25 – 📊 Real-time merchandising optimization and ad performance
21:01 – ⚙️ Letting AI handle repeatable, data-driven tasks
21:40 – 🙌 Final thoughts and future of AI in automotive
Connect with Sanjay Varnwal at https://www.linkedin.com/in/sanjaykv/
Learn more about Spyne at https://www.spyne.ai
⭐️ Love the podcast? Please leave us a review here — even one sentence helps! Consider including your LinkedIn or Instagram handle so we can thank you personally!
We have a daily email!
✉️ Sign up for our free and fun-to-read daily email for a quick shot of relevant news in automotive retail, media, and pop culture.
🎧 Like and follow our other podcasts:
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Paul J Daly (00:00):
Michael, you have some new glasses. They talk to
me.
Unknown (00:09):
This is Auto Collabs.
Paul J Daly (00:11):
I mean, I thought, I thought, Well,
Kyle Mountsier (00:15):
Michael, let me break it to you. Those are just
the voices in your head. They'vealways been there. That's just
what
Paul J Daly (00:21):
they did to make you feel better about it. It's a
new therapy. They give youglasses. They say, talk to you
so that you think it's, oh, I'mnormal, but you're not. I
actually thought you were tryingout to be the new lead singer,
Weezer.
Michael Cirillo (00:34):
That would be, man, that you just brought me
right back to the seventh grade.
Paul J Daly (00:39):
I mean, I was pretty much trade and I would,
yeah, I mean, you know, guys, ifWeezer calls me one day you're
gone is like, say, hey, we needa new front man, or we need a
drummer. Like, you're just nevergonna see me again. I'm just
Michael Cirillo (00:55):
pretty much any band. Like, there's, I have,
really any band. You
Paul J Daly (00:58):
trade us for, any band.
Kyle Mountsier (01:00):
Well, he's like some death metal van band. Done
Michael Cirillo (01:05):
these shorts, Jonas Brothers. Now the audience
is throwing me under the bus,bro, I don't think I'm not
missing this. You got the betterbids? They'd rather be your
Paul J Daly (01:16):
friend. Listen, listen, it's
Michael Cirillo (01:20):
just exposure.
When I'd take any band, at thispoint, I would take like, I'd
join the wiggles.
Paul J Daly (01:28):
Go hard with the kid. Strangely enough, that is
probably the most likely band tocall you, because you would make
an amazing wiggle,
Michael Cirillo (01:36):
just amazing.
I'd be the hype guy, like themighty, mighty boss tones. I
wear a gray suit in the
Paul J Daly (01:41):
back with with Jack Black vibes, and you're dancing,
Unknown (01:46):
just kicking everywhere.
Kyle Mountsier (01:48):
Well, all that has nothing to do with the fact
that we are hanging out withSanjay Vaughn, wall from from
spine, AI, talking AI. Man,we're talking AI. We're gonna be
way more intelligent, hopefully,than us talking about the
wiggles. But if you want tolisten to the wiggles,
Paul J Daly (02:08):
no, he's wherever you want to, Sanjay is easily
going to be bring the collectiveIQ average up by many points
once he enters this room.
Kyle Mountsier (02:16):
Yes. Why? Look, we really hope you enjoy this
conversation with Sanjay.
Paul J Daly (02:25):
Sanjay, how are you today? Thank you so much for
joining us.
Sanjay Varnwal (02:28):
Hey, thanks for inviting me. For the for the for
your show. Paul Kyle andMichael,
Paul J Daly (02:34):
yeah. So, so I'm sure to talk a lot about AI
today, but I, I think this isgoing to be the question I start
asking people, whenever we'retalking about an AI topic, tell
me about the first experienceyou ever had with an AI product.
Sanjay Varnwal (02:48):
Oh, that's tough. I mean, no one has asked
me this question before.
Unknown (02:52):
Yeah,
Sanjay Varnwal (02:55):
nailed it. So I'm remembering few of the
things I used to work at Amazonin turn, 14 or 15 or so. And
there, I used to travel toSeattle often, and I saw Amazon
doing a pilot of a store calledAmazon Go at that point of time.
And it was one of its kind. Whenit was launched. You can just
(03:17):
check and go inside the store.
You don't need to do anything.
Just pick up goods from thereand go out and everything is
like, checked out for you,payment is debited, everything.
So Amazon was doing thatparticular pilot. Of course, the
smaller AI use cases I wouldhave seen in the past, but that
was the first big, working,practical use case that I saw on
(03:38):
the ground,
Kyle Mountsier (03:39):
which, yeah, I think that that's interesting,
because for the majority of theworld, AI didn't really exist
until, I don't know 2022 intheir brains, right? But these
big corporations have beendealing with AI since, like,
really, the early 2000s therewere already inklings of like,
how do we create things that cando stuff on its own, right?
Sanjay Varnwal (04:02):
Absolutely, although Google recommendations
and all these things like, thereis an element of AI behind it
already, right? So, go
Paul J Daly (04:10):
ahead. Go ahead, continue. Yeah. I mean
Sanjay Varnwal (04:13):
the AI in terms of data analysis, data
analytics, some of these techhas been existing for long, long
time, but computer vision, whereyou see, like the magic
happening in front of you issomething that that started
getting into prominence in thelast like seven, eight years or
so, yeah, just blown it up.
Paul J Daly (04:31):
When was that movie that came out with that, that
little kid, I just think of himas like the sad little boy,
because he's got those sad eyes.
Haley, Joel Osmond is and themovie was called AI, Oh, yeah.
And I feel like that was thefirst time that I even, like,
thought about artificialintelligence, yeah, right. And
that was probably, I mean, Igotta say that that movie had to
come out 15 to 20 years ago. Butif that movie came out, I bet, I
(04:53):
bet there was something going onin the tech world that somebody
was like, You know what? We seethis coming. We're gonna make. A
movie about it, and now it'slike movies reality.
Michael Cirillo (05:02):
I love guys like Neil, degrasse, Tyson and
stuff, though, because theybring everything back to its
most rudimentary form. They'relike, you know him? He's like,
Oh, you want flying cars. Wehave those. They're called
helicopters. And Sanjay, youmade me think of this because,
you know, with Google, with whatGoogle's worked on over the
(05:23):
last, I don't know, 20 years,anybody that's, if you want
light reading, go learn aboutGoogle's Rank Brain, because
that's been around for quitesome time, and that is, I mean,
that is what Skynet might bemade of.
Sanjay Varnwal (05:38):
Rank Brain, fundamentally, is the core
machine learning base on theirsearch algorithms are built,
right? So I haven't studied indepth, but I've heard that this
is, this is the base basedtechnology on which the entire
Google search functionsglobally. So got it,
Michael Cirillo (05:58):
and it's like super it's like the treasury of
the United States of America,like it is ironclad. There are,
there are like, Rank Brain,Google, Rank Brain offices all
over the world that you know.
For example, we heard throughthe grapevine that there's an
office in a lab at theUniversity of Alberta in
Edmonton, but there's no Googlesignage, any like they don't
(06:20):
want people to know what's goingon.
Paul J Daly (06:25):
I see, yeah,
Michael Cirillo (06:28):
it's deep, you know? It's funny, Sanjay, like
we were just talking. I saw thisnotification in Slack that
slacks just released. Its AI toour to our organization, and so
I started playing around withit. I'm curious. You know what
your take is on, on, on this ifAI, let me back up see my
(06:50):
brain's moving so fast, maybethis is I need the AI.
Paul J Daly (06:53):
It's those meta Ray Bans. You the questions it's
asking.
Michael Cirillo (06:56):
Yeah, John Stina is in my ear telling me
what to ask. I saw a memerecently, and it said in the
year, what was it in the year2050, that this is how the
generation will communicate? Andhe was basically like a caveman
again, prompting GPT. He waslike, Jeremy.
Unknown (07:18):
Good thing to say to girl, you know, like, there's
like that just morphed into GPTto the degree that we can't even
communicate as humans anymore.
What are you seeing? I mean, Iknow you're heavy into this
space and software developmentand the implications and
applications of AI. Where do yousee? What's happening out there?
What's going on?
Sanjay Varnwal (07:40):
I think you said it right. I'm not too sure if
that would be the situation, butpotentially could be. But yeah,
as you see things around theway, the behaviors are changing
for all of us, right? So I mean,earlier, we used to think, and
then you used to write or usedto say anything, but now I see
(08:01):
myself, my behavior is changingto be whenever I need to respond
to anyone, I I ask chat. GPT,hey, this is the profile. This
is me. This is the context. Canyou come up with this direct, I
mean very I mean the kind ofresponse that that elicits some
response from them, right? So,and it comes out with very well
(08:22):
researched answer or theresponses. So they're like, just
one of the use cases. I keepusing chat, GPT for a variety of
use cases, and have stoppedapplying my brains to a lot of
work. I think that, you
Kyle Mountsier (08:36):
know, I think that there's different levels of
usage. One is like, here's ahere's an email. Give me a
response, but you actually aredoing some critical thinking in
that thing that you just said,which is, like, here's, here's
something that someone said,here's a little bit more about
them, here's a little bit moreabout me. What I want is this,
this, this and this, give mesome feedback in order to do
(08:59):
that. And I think that that's,that's the layer that I think
that we have to be thinkingabout any of these tools with,
is, how do I take my like,knowledge of the world, my
knowledge of the context of thesituation, inject into what I'm
what I'm using, and put it inI'd like to kind of pivot the
conversation into, you have aunique knowledge. You worked at
(09:22):
Amazon, have worked at some ofthese larger companies, and you
saw it fit to move intoautomotive as a vertical that
you wanted to serve. Why wasthat the layer that, like you,
you brought to the conversationthat was, hey, I'm seeing all
these things move in the world,and I'm looking at retail auto
and saying that's where I needto see a fit,
Sanjay Varnwal (09:45):
yes, yes. So we picked up merchandising as a
core problem in the industry,right? So, amen
Paul J Daly (09:51):
say that twice.
Sanjay Varnwal (09:55):
So mostly like digital merchandising, it was
the problem that I picked up in.
The in this particular industry,working at Amazon and so many
other e commerce firms, realizethat sellers, day in day out,
struggle with producing highquality merchandising. I mean,
the bigger sellers can hireexpensive photographers
agencies, but the smaller ones,long tail, do not do that. And
the photo qualities and the ingeneral media quality is shit
(10:19):
very bad, it's and that also wasnot helping them sell more
products online. So the ideawas, can we create something
which is more AI first in itsapproach, understands what
product has to be shot, how toshoot it like a photographer,
and then process those imagesinto something that Amazon
understands, that maybe AutoTrader understand that. So
these, like larger platforms,they have given the guidelines
(10:42):
and the course shoots areattuned for those guidelines.
Picked up. This worked in thisidea, launched something, ran it
for two years, and then figuredout that if you, if you are
going to do everything, will beaverage at everything, right? So
real estate has to be shotdifferently, versus fashion,
model versus the product, versusthe car versus other other
(11:03):
objects, right? And that's wherewe realized that, okay, we need
to pick one category. And carwas always a passion. I knew a
lot of dealers in my locality,so assured them the product.
They were like, very happy. Wegot it rolled out to some of the
biggest companies in thecountry, in India, and then we
saw that this use case wasworking like phenomenally well,
(11:23):
and US Europe being extremelyhuge market. We thought, Okay,
this is, this is the directionthat we led to. So
Kyle Mountsier (11:30):
you actually went at, like, just general
merchandising. Could have beenany type of product. First, is
that what you're saying?
Sanjay Varnwal (11:36):
Yeah, that we focused on energy to build,
like, the best product inautomotive? In
Kyle Mountsier (11:41):
auto? Yeah, I think that that's, that's an
interesting start and pivot.
Because when you do look atAmazon and you look at these
large third party aggregatorsthat are that are very similar
to auto, I I always, I alwaysmake this argument, like,
everyone's like, I don't want tobe on the third parties. And
it's like, well, the rest of theworld operates that way. So why
wouldn't we too? I mean, anybodylike Amazon, right? That's a
(12:01):
third party aggregator of theproducts, right? But when you
think about, you know, theglobal market, the global
merchandising efforts, you know,Nike spends millions and
millions and millions of dollarsa year to merchandise their
shopping results pages, right?
Because they know that that nextclick is critical, right? And,
(12:23):
and we kind of like, you know,wait 11 days till we get a car
photod and online, you know,because it has to get through
the shop or whatever, and then,and then it's maybe got photos.
And we've wasted 11 days. And Ithink that that merchandising
flow in auto, you know,obviously we have unique VINs.
It's a little bit differentthan, like, one shoe that you
(12:44):
can shoot one time and haveavailable for six months. So
it's a different type ofproblem. But I love that you're
trying to solve it. What havebeen some of the hardest things
that you've encountered alongthe way, when you're like, well,
we set out to build it this way,but that was actually harder, or
that that was a mistake or athing that you didn't expect
along the way.
Sanjay Varnwal (13:05):
Yeah. I mean, the goal that we with, which we
started was we saw Carmaxmerchandising in very depth, car
max or Carvana merchandising invery depth, right? And
understood that some of thelargest companies are setting up
those 100,000 worth of studiosin every single location to do
their photo shoot. And the ideawas that, hey, small dealer
(13:27):
would not be setting up that100,000 worth of booth because
of real estate constraint andmaintenance constraint and
initial investment, right? Butthey all want that kind of a
catalog merchandisingexperience. Idea was, can we get
every single dealer to thatparticular experience without
having to invest in that studio?
That was the idea that, I mean,it should be indifferentiable
(13:50):
from human eye, whether the carwas actually shot in the studio
or shot outside. That was thebenchmark that we were tracing
and we launched the product. We,I mean, you can check some of
these studios. These are veryrealistic, glassy kind of
reflections of the car. When,when we produce outcomes, right?
And so we, we create behind thescenes some 7080, computer
(14:14):
vision models, focusing onreflections, focus focusing on
stabilization, focusing oninterior view generation,
focusing on like, lot of theseproblems we understand, like,
hundreds of parts of the car,and every part are tuned in a
different fashion when weprocess the image for this car.
So the more hard problem that wegot into was, one was
(14:35):
stabilization, because when youshoot a car and move around it,
your hands are unstable like andyou when you try to place it in
the studio, your car will lookwobbly because the studio is
perfect. How do you make thatcar spin perfect in that
particular studio? So that wasone of the technologies that we
worked almost like over a yearto perfect with. Now, if you see
(14:55):
our 360 spins, no matter how youshoot it. Outside, it will
always be perfectly placed inthat particular studio, very,
very smooth. This is oneproblem. The second was lighting
problem, right? So exposure,lighting, some of these, oh my
gosh, yes. If you shoot in theopen yard, like harsh sunlight
or during the later half of theday, so you'll see all like one
side is lit, other other side isdark. Some of these problems we
(15:20):
kept on encountering. And so wewe created those exposure
correction, Color Correctionalgorithms that even the light
on the car, and based on thestudio's outlook, it will, it
will, it will project that kindof a tent on the on that
particular car, so it will looklike that it was shot in the
studio. Some of these things wedid which, which were, like,
(15:40):
really hard problems. Like, westayed on these problems for
months, six months to 12.
Kyle Mountsier (15:46):
How long did it take you to make sure that
mirrors didn't disappear?
Because that was my favoritething in, like, early even, like
two, three years ago, like, it'sjust missing a whole mirror,
right? Or, like, the tires aregone. It's like, who lifted this
car and made it a hover car?
Back to the hover car thing,right? Yeah. Like, why? Why was
(16:08):
that an issue, and how did yousolve it?
Sanjay Varnwal (16:10):
Yeah, so. So, whenever you create any kind of
computer vision model or anykind of AI model, it has to
understand the shape and size ofthat object on which it is
operating, right? So that isalso one of the reasons why we
went into automotive, becausethe shapes were defined earlier
we were doing like multiplecategories. So I mean, a human
versus a bag versus anything,right? So everything has a
(16:33):
different shape. So when wepicked up cards, we trained our
algorithms with, I mean, fewmillions of images which were
custom trained by our ownannotators, and we specifically
focused on this problem likemirror was one of the biggest,
hardest problems that we solved,mirrors and antennas. So yeah,
the antennas for sure, theantennas, right? So these two
(16:54):
were one of the toughestproblems that we solved in
background removal, as well asthose open, open door shots that
you take right so when, whenthings can go wrong. So some of
these, like specific hardwareuse cases, we generated our own
custom data in like hundreds of1000s, and then trained our
models to make that to work.
Now. Now it works with like 9899% plus accuracy. And you give
it any kind of image, it will
Paul J Daly (17:18):
work. I mean, you all do? You all do so much more
than photos. You know, I heardthat you're building like an AI
powered teammate, and I'massuming this kind of, like,
delves into the whole realm ofagentic AI and so. So tell us
what you're excited about. Like,what does the roadmap look for
look like for what you'rebuilding right now?
Sanjay Varnwal (17:40):
So, so we are.
We are. I mean visualizing us asa company which will build AI
first experiences for automotiveindustry, right? So
merchandising was one of the usecases that we targeted. And
merchandising natural extensionis to go in other use cases of
inventory, which is, how do youprice, how do you inspect? How
do you syndicate the car? Well,so this is like one, one
(18:02):
spectrum on which will go verydeep and build AI native
experiences. The second, secondpart where we are now getting
into is the this entire customerinteraction of the dealers,
right, using agent AI. Sowhatever we are doing now, we
are looking to bring in agentAI, first behavior. So we are
building like bots for sales,service, finance, and these
(18:23):
could be like voice chat,emails, everything, and we
deployed in like variety of usecases. So you are already doing
like a number of pilots withdealers here and seeing like
amazing, amazing results.
Kyle Mountsier (18:40):
Explain the agentic, AI, just a little bit,
because I think some people getlost with that, like, explain
exactly what it's doing and whatyou're hoping to have it
accomplished. Because when wetalk about AI, a lot of people
are talking about, like, theability to converse with it, but
you're talking about takingaction. So go into like, what's
like in a specific use case,what are you what are you seeing
(19:01):
the agent, agentic. Ai, doingso.
Sanjay Varnwal (19:04):
Agentic. Ai, I mean conceptual level, it is
like you do set of things andthat that lead to certain
results happening, right? AgentA is just automating all of
these things, where, instead ofa human, that particular agent
is doing things and thenpresenting you with the results
that, hey, I have done this.
These are the results. I mean,if you want to set auto rule
(19:27):
that, okay, if these results areabove 90% confidence, just go
ahead execute them. Or I needto, like, check with it and do
the action that, whether I needto take it or or reject it,
right? So, Agent decay ingeneral, is automating lot of
stuff that you do. Instead ofyou working on the software, the
software is working for you, andit is telling you the options
(19:50):
that, okay, I've done this. Doyou want to present it? Do you
want to publish it? Even thatcard could be automated, right?
So, merchandising, we are a.
Doing a lot of agent again, likewe did the shot, instead of
users signing off, we are, wehave we are. We are basically
signing off all of these things.
We are changing the cars imagesin real time on website, without
informing the dealer, right,based on which image is doing
(20:13):
better. So some of these, thesethings, we are, we are
automating on its own, becausewe are controlling the website
infrastructure now for media andsimilarly here. So some of these
things, right?
Kyle Mountsier (20:25):
So like, which, which image should be the first
image, or something like that,
Sanjay Varnwal (20:29):
one of the first use cases, right? When you get
into pricing, you will be ableto when you manage the ads.
Let's say there is an agent, awhich is, which is AD, which is
contextualized or trained on theads behavior, so it
automatically will will increaseor decrease the budgets or
choose the mediums where itneeds to spend more based on how
(20:50):
the clicks are performing,right? So you don't need a
person to do all these thingsfor you. The agent Aki, which is
trained on those use cases, willdo it for you, and it is way
more efficient.
Kyle Mountsier (21:01):
Yeah, I think, I think that the name of the game
with AI is efficiency, and it's,how do I take the tasks that can
be repeated, understood when itcomes to data, and hand it off
to something that can do itfaster and more efficiently, to
free my people up to do thething that they're best at?
Sanjay, thank you so much forspending time with us today. We
(21:22):
We walked the gamut from like,what you saw on Amazon and what
we saw in the early 2000s tonow, what what you're doing and
what you're what you'reinnovating on top of and I can't
wait to see what you and spinedo as you deeper dive into
automotive thanks for joining ustoday on Auto Collabs.
Sanjay Varnwal (21:40):
Yeah, thanks a lot for having me here, Kyle,
Paul and Mike. Great having chatwith you. You guys are like huge
father.
Paul J Daly (21:52):
One of the things I'm super impressed with is the
hyper focus of Sanjay and histeam to fix small problems, and
how all of those things add upto fixing big problems.
Literally, that sounds maddeningto me, focusing that long on the
problem of reflections, but it'sindicative of the fact that they
understand that the small thingsimpact the big things. And we
(22:14):
see that a lot, even as wedeploy AI, quite a bit within
our business. If you don't focuson the small things, the big
things end up being terrible.
Michael Cirillo (22:22):
This is, like, the worst part of our human
nature, which is that we want toperpetually be at the top step
already. Yeah, we never want toclimb the steps to get there,
not realizing that every stepgives us a lesson needed to
actually stand at the top step.
But like, you want a real world,very painful implication of why
people need to focus on thesmaller thing and be great at
it. Look at that dude. I justwatched that documentary about
(22:45):
the guy the deep blue ocean,whatever
Paul J Daly (22:50):
the oh, he went down and like that. The whole
warnings
Michael Cirillo (22:54):
of like, you need to focus on that thing and
make that thing better. And nextthing you know, it's like,
implodes in the middle of theocean, right? And, and we see
that, and we're like, what anidiot. But we're doing these
things in our businesses everyday. We're ignoring, you know,
(23:14):
the small thing. And so to yourpoint about that's very
inspiring to have thediscipline. And
Kyle Mountsier (23:19):
I'm asking chat GPT, how do I change the world
today, right? It's right. How doI make sure that this employee
is able to deploy this specifictask and function well within my
organization, those type ofspecific questions like, How do
I make sure that mirrors show upevery time I do a background
(23:42):
replace. You know, that's howspecific we have to be in
everything, and especially whenyou're deploying AI tools across
your organization. Doesn'tmatter if you're like doing
communicative, communicative AI,agentic AI, you know, photo
backgrounding, whatever it maybe, being very specific with the
task that that person ortechnology is accomplishing. Is
(24:05):
it going to be extremelyimportant as we move forward,
Paul J Daly (24:08):
fewer better, there you go. Well, thank you to our
friend Sanjay and the team atspine for putting their heart
soul and all the intention tomaking this industry better. On
behalf of Kyle Mountsier,Michael Cirillo. Michael
Cirillo, new Ray Ban, metaglasses and myself. Thank you so
much for joining us on AutoCollabs.
Unknown (24:27):
Sign up for our free and fun to read daily email for
a free shot of relevant news andautomotive retail media and pop
culture. You can get itnow@asotu.com that's asotu.com
if you love this podcast, pleaseleave us a review and share it
with a friend. Thanks again forlistening. We'll see you next
time.
Paul J Daly (24:58):
Welcome to autocala. US,
Unknown (25:03):
recording you.
Michael, you have some new glasses. They talk to
me.
Unknown (00:09):
This is Auto Collabs.
Paul J Daly (00:11):
I mean, I thought, I thought, Well,
Kyle Mountsier (00:15):
Michael, let me break it to you. Those are just
the voices in your head. They'vealways been there. That's just
what
Paul J Daly (00:21):
they did to make you feel better about it. It's a
new therapy. They give youglasses. They say, talk to you
so that you think it's, oh, I'mnormal, but you're not. I
actually thought you were tryingout to be the new lead singer,
Weezer.
Michael Cirillo (00:34):
That would be, man, that you just brought me
right back to the seventh grade.
Paul J Daly (00:39):
I mean, I was pretty much trade and I would,
yeah, I mean, you know, guys, ifWeezer calls me one day you're
gone is like, say, hey, we needa new front man, or we need a
drummer. Like, you're just nevergonna see me again. I'm just
Michael Cirillo (00:55):
pretty much any band. Like, there's, I have,
really any band. You
Paul J Daly (00:58):
trade us for, any band.
Kyle Mountsier (01:00):
Well, he's like some death metal van band. Done
Michael Cirillo (01:05):
these shorts, Jonas Brothers. Now the audience
is throwing me under the bus,bro, I don't think I'm not
missing this. You got the betterbids? They'd rather be your
Paul J Daly (01:16):
friend. Listen, listen, it's
Michael Cirillo (01:20):
just exposure.
When I'd take any band, at thispoint, I would take like, I'd
join the wiggles.
Paul J Daly (01:28):
Go hard with the kid. Strangely enough, that is
probably the most likely band tocall you, because you would make
an amazing wiggle,
Michael Cirillo (01:36):
just amazing.
I'd be the hype guy, like themighty, mighty boss tones. I
wear a gray suit in the
Paul J Daly (01:41):
back with with Jack Black vibes, and you're dancing,
Unknown (01:46):
just kicking everywhere.
Kyle Mountsier (01:48):
Well, all that has nothing to do with the fact
that we are hanging out withSanjay Vaughn, wall from from
spine, AI, talking AI. Man,we're talking AI. We're gonna be
way more intelligent, hopefully,than us talking about the
wiggles. But if you want tolisten to the wiggles,
Paul J Daly (02:08):
no, he's wherever you want to, Sanjay is easily
going to be bring the collectiveIQ average up by many points
once he enters this room.
Kyle Mountsier (02:16):
Yes. Why? Look, we really hope you enjoy this
conversation with Sanjay.
Paul J Daly (02:25):
Sanjay, how are you today? Thank you so much for
joining us.
Sanjay Varnwal (02:28):
Hey, thanks for inviting me. For the for the for
your show. Paul Kyle andMichael,
Paul J Daly (02:34):
yeah. So, so I'm sure to talk a lot about AI
today, but I, I think this isgoing to be the question I start
asking people, whenever we'retalking about an AI topic, tell
me about the first experienceyou ever had with an AI product.
Sanjay Varnwal (02:48):
Oh, that's tough. I mean, no one has asked
me this question before.
Unknown (02:52):
Yeah,
Sanjay Varnwal (02:55):
nailed it. So I'm remembering few of the
things I used to work at Amazonin turn, 14 or 15 or so. And
there, I used to travel toSeattle often, and I saw Amazon
doing a pilot of a store calledAmazon Go at that point of time.
And it was one of its kind. Whenit was launched. You can just
(03:17):
check and go inside the store.
You don't need to do anything.
Just pick up goods from thereand go out and everything is
like, checked out for you,payment is debited, everything.
So Amazon was doing thatparticular pilot. Of course, the
smaller AI use cases I wouldhave seen in the past, but that
was the first big, working,practical use case that I saw on
(03:38):
the ground,
Kyle Mountsier (03:39):
which, yeah, I think that that's interesting,
because for the majority of theworld, AI didn't really exist
until, I don't know 2022 intheir brains, right? But these
big corporations have beendealing with AI since, like,
really, the early 2000s therewere already inklings of like,
how do we create things that cando stuff on its own, right?
Sanjay Varnwal (04:02):
Absolutely, although Google recommendations
and all these things like, thereis an element of AI behind it
already, right? So, go
Paul J Daly (04:10):
ahead. Go ahead, continue. Yeah. I mean
Sanjay Varnwal (04:13):
the AI in terms of data analysis, data
analytics, some of these techhas been existing for long, long
time, but computer vision, whereyou see, like the magic
happening in front of you issomething that that started
getting into prominence in thelast like seven, eight years or
so, yeah, just blown it up.
Paul J Daly (04:31):
When was that movie that came out with that, that
little kid, I just think of himas like the sad little boy,
because he's got those sad eyes.
Haley, Joel Osmond is and themovie was called AI, Oh, yeah.
And I feel like that was thefirst time that I even, like,
thought about artificialintelligence, yeah, right. And
that was probably, I mean, Igotta say that that movie had to
come out 15 to 20 years ago. Butif that movie came out, I bet, I
(04:53):
bet there was something going onin the tech world that somebody
was like, You know what? We seethis coming. We're gonna make. A
movie about it, and now it'slike movies reality.
Michael Cirillo (05:02):
I love guys like Neil, degrasse, Tyson and
stuff, though, because theybring everything back to its
most rudimentary form. They'relike, you know him? He's like,
Oh, you want flying cars. Wehave those. They're called
helicopters. And Sanjay, youmade me think of this because,
you know, with Google, with whatGoogle's worked on over the
(05:23):
last, I don't know, 20 years,anybody that's, if you want
light reading, go learn aboutGoogle's Rank Brain, because
that's been around for quitesome time, and that is, I mean,
that is what Skynet might bemade of.
Sanjay Varnwal (05:38):
Rank Brain, fundamentally, is the core
machine learning base on theirsearch algorithms are built,
right? So I haven't studied indepth, but I've heard that this
is, this is the base basedtechnology on which the entire
Google search functionsglobally. So got it,
Michael Cirillo (05:58):
and it's like super it's like the treasury of
the United States of America,like it is ironclad. There are,
there are like, Rank Brain,Google, Rank Brain offices all
over the world that you know.
For example, we heard throughthe grapevine that there's an
office in a lab at theUniversity of Alberta in
Edmonton, but there's no Googlesignage, any like they don't
(06:20):
want people to know what's goingon.
Paul J Daly (06:25):
I see, yeah,
Michael Cirillo (06:28):
it's deep, you know? It's funny, Sanjay, like
we were just talking. I saw thisnotification in Slack that
slacks just released. Its AI toour to our organization, and so
I started playing around withit. I'm curious. You know what
your take is on, on, on this ifAI, let me back up see my
(06:50):
brain's moving so fast, maybethis is I need the AI.
Paul J Daly (06:53):
It's those meta Ray Bans. You the questions it's
asking.
Michael Cirillo (06:56):
Yeah, John Stina is in my ear telling me
what to ask. I saw a memerecently, and it said in the
year, what was it in the year2050, that this is how the
generation will communicate? Andhe was basically like a caveman
again, prompting GPT. He waslike, Jeremy.
Unknown (07:18):
Good thing to say to girl, you know, like, there's
like that just morphed into GPTto the degree that we can't even
communicate as humans anymore.
What are you seeing? I mean, Iknow you're heavy into this
space and software developmentand the implications and
applications of AI. Where do yousee? What's happening out there?
What's going on?
Sanjay Varnwal (07:40):
I think you said it right. I'm not too sure if
that would be the situation, butpotentially could be. But yeah,
as you see things around theway, the behaviors are changing
for all of us, right? So I mean,earlier, we used to think, and
then you used to write or usedto say anything, but now I see
(08:01):
myself, my behavior is changingto be whenever I need to respond
to anyone, I I ask chat. GPT,hey, this is the profile. This
is me. This is the context. Canyou come up with this direct, I
mean very I mean the kind ofresponse that that elicits some
response from them, right? So,and it comes out with very well
(08:22):
researched answer or theresponses. So they're like, just
one of the use cases. I keepusing chat, GPT for a variety of
use cases, and have stoppedapplying my brains to a lot of
work. I think that, you
Kyle Mountsier (08:36):
know, I think that there's different levels of
usage. One is like, here's ahere's an email. Give me a
response, but you actually aredoing some critical thinking in
that thing that you just said,which is, like, here's, here's
something that someone said,here's a little bit more about
them, here's a little bit moreabout me. What I want is this,
this, this and this, give mesome feedback in order to do
(08:59):
that. And I think that that's,that's the layer that I think
that we have to be thinkingabout any of these tools with,
is, how do I take my like,knowledge of the world, my
knowledge of the context of thesituation, inject into what I'm
what I'm using, and put it inI'd like to kind of pivot the
conversation into, you have aunique knowledge. You worked at
(09:22):
Amazon, have worked at some ofthese larger companies, and you
saw it fit to move intoautomotive as a vertical that
you wanted to serve. Why wasthat the layer that, like you,
you brought to the conversationthat was, hey, I'm seeing all
these things move in the world,and I'm looking at retail auto
and saying that's where I needto see a fit,
Sanjay Varnwal (09:45):
yes, yes. So we picked up merchandising as a
core problem in the industry,right? So, amen
Paul J Daly (09:51):
say that twice.
Sanjay Varnwal (09:55):
So mostly like digital merchandising, it was
the problem that I picked up in.
The in this particular industry,working at Amazon and so many
other e commerce firms, realizethat sellers, day in day out,
struggle with producing highquality merchandising. I mean,
the bigger sellers can hireexpensive photographers
agencies, but the smaller ones,long tail, do not do that. And
the photo qualities and the ingeneral media quality is shit
(10:19):
very bad, it's and that also wasnot helping them sell more
products online. So the ideawas, can we create something
which is more AI first in itsapproach, understands what
product has to be shot, how toshoot it like a photographer,
and then process those imagesinto something that Amazon
understands, that maybe AutoTrader understand that. So
these, like larger platforms,they have given the guidelines
(10:42):
and the course shoots areattuned for those guidelines.
Picked up. This worked in thisidea, launched something, ran it
for two years, and then figuredout that if you, if you are
going to do everything, will beaverage at everything, right? So
real estate has to be shotdifferently, versus fashion,
model versus the product, versusthe car versus other other
(11:03):
objects, right? And that's wherewe realized that, okay, we need
to pick one category. And carwas always a passion. I knew a
lot of dealers in my locality,so assured them the product.
They were like, very happy. Wegot it rolled out to some of the
biggest companies in thecountry, in India, and then we
saw that this use case wasworking like phenomenally well,
(11:23):
and US Europe being extremelyhuge market. We thought, Okay,
this is, this is the directionthat we led to. So
Kyle Mountsier (11:30):
you actually went at, like, just general
merchandising. Could have beenany type of product. First, is
that what you're saying?
Sanjay Varnwal (11:36):
Yeah, that we focused on energy to build,
like, the best product inautomotive? In
Kyle Mountsier (11:41):
auto? Yeah, I think that that's, that's an
interesting start and pivot.
Because when you do look atAmazon and you look at these
large third party aggregatorsthat are that are very similar
to auto, I I always, I alwaysmake this argument, like,
everyone's like, I don't want tobe on the third parties. And
it's like, well, the rest of theworld operates that way. So why
wouldn't we too? I mean, anybodylike Amazon, right? That's a
(12:01):
third party aggregator of theproducts, right? But when you
think about, you know, theglobal market, the global
merchandising efforts, you know,Nike spends millions and
millions and millions of dollarsa year to merchandise their
shopping results pages, right?
Because they know that that nextclick is critical, right? And,
(12:23):
and we kind of like, you know,wait 11 days till we get a car
photod and online, you know,because it has to get through
the shop or whatever, and then,and then it's maybe got photos.
And we've wasted 11 days. And Ithink that that merchandising
flow in auto, you know,obviously we have unique VINs.
It's a little bit differentthan, like, one shoe that you
(12:44):
can shoot one time and haveavailable for six months. So
it's a different type ofproblem. But I love that you're
trying to solve it. What havebeen some of the hardest things
that you've encountered alongthe way, when you're like, well,
we set out to build it this way,but that was actually harder, or
that that was a mistake or athing that you didn't expect
along the way.
Sanjay Varnwal (13:05):
Yeah. I mean, the goal that we with, which we
started was we saw Carmaxmerchandising in very depth, car
max or Carvana merchandising invery depth, right? And
understood that some of thelargest companies are setting up
those 100,000 worth of studiosin every single location to do
their photo shoot. And the ideawas that, hey, small dealer
(13:27):
would not be setting up that100,000 worth of booth because
of real estate constraint andmaintenance constraint and
initial investment, right? Butthey all want that kind of a
catalog merchandisingexperience. Idea was, can we get
every single dealer to thatparticular experience without
having to invest in that studio?
That was the idea that, I mean,it should be indifferentiable
(13:50):
from human eye, whether the carwas actually shot in the studio
or shot outside. That was thebenchmark that we were tracing
and we launched the product. We,I mean, you can check some of
these studios. These are veryrealistic, glassy kind of
reflections of the car. When,when we produce outcomes, right?
And so we, we create behind thescenes some 7080, computer
(14:14):
vision models, focusing onreflections, focus focusing on
stabilization, focusing oninterior view generation,
focusing on like, lot of theseproblems we understand, like,
hundreds of parts of the car,and every part are tuned in a
different fashion when weprocess the image for this car.
So the more hard problem that wegot into was, one was
(14:35):
stabilization, because when youshoot a car and move around it,
your hands are unstable like andyou when you try to place it in
the studio, your car will lookwobbly because the studio is
perfect. How do you make thatcar spin perfect in that
particular studio? So that wasone of the technologies that we
worked almost like over a yearto perfect with. Now, if you see
(14:55):
our 360 spins, no matter how youshoot it. Outside, it will
always be perfectly placed inthat particular studio, very,
very smooth. This is oneproblem. The second was lighting
problem, right? So exposure,lighting, some of these, oh my
gosh, yes. If you shoot in theopen yard, like harsh sunlight
or during the later half of theday, so you'll see all like one
side is lit, other other side isdark. Some of these problems we
(15:20):
kept on encountering. And so wewe created those exposure
correction, Color Correctionalgorithms that even the light
on the car, and based on thestudio's outlook, it will, it
will, it will project that kindof a tent on the on that
particular car, so it will looklike that it was shot in the
studio. Some of these things wedid which, which were, like,
(15:40):
really hard problems. Like, westayed on these problems for
months, six months to 12.
Kyle Mountsier (15:46):
How long did it take you to make sure that
mirrors didn't disappear?
Because that was my favoritething in, like, early even, like
two, three years ago, like, it'sjust missing a whole mirror,
right? Or, like, the tires aregone. It's like, who lifted this
car and made it a hover car?
Back to the hover car thing,right? Yeah. Like, why? Why was
(16:08):
that an issue, and how did yousolve it?
Sanjay Varnwal (16:10):
Yeah, so. So, whenever you create any kind of
computer vision model or anykind of AI model, it has to
understand the shape and size ofthat object on which it is
operating, right? So that isalso one of the reasons why we
went into automotive, becausethe shapes were defined earlier
we were doing like multiplecategories. So I mean, a human
versus a bag versus anything,right? So everything has a
(16:33):
different shape. So when wepicked up cards, we trained our
algorithms with, I mean, fewmillions of images which were
custom trained by our ownannotators, and we specifically
focused on this problem likemirror was one of the biggest,
hardest problems that we solved,mirrors and antennas. So yeah,
the antennas for sure, theantennas, right? So these two
(16:54):
were one of the toughestproblems that we solved in
background removal, as well asthose open, open door shots that
you take right so when, whenthings can go wrong. So some of
these, like specific hardwareuse cases, we generated our own
custom data in like hundreds of1000s, and then trained our
models to make that to work.
Now. Now it works with like 9899% plus accuracy. And you give
it any kind of image, it will
Paul J Daly (17:18):
work. I mean, you all do? You all do so much more
than photos. You know, I heardthat you're building like an AI
powered teammate, and I'massuming this kind of, like,
delves into the whole realm ofagentic AI and so. So tell us
what you're excited about. Like,what does the roadmap look for
look like for what you'rebuilding right now?
Sanjay Varnwal (17:40):
So, so we are.
We are. I mean visualizing us asa company which will build AI
first experiences for automotiveindustry, right? So
merchandising was one of the usecases that we targeted. And
merchandising natural extensionis to go in other use cases of
inventory, which is, how do youprice, how do you inspect? How
do you syndicate the car? Well,so this is like one, one
(18:02):
spectrum on which will go verydeep and build AI native
experiences. The second, secondpart where we are now getting
into is the this entire customerinteraction of the dealers,
right, using agent AI. Sowhatever we are doing now, we
are looking to bring in agentAI, first behavior. So we are
building like bots for sales,service, finance, and these
(18:23):
could be like voice chat,emails, everything, and we
deployed in like variety of usecases. So you are already doing
like a number of pilots withdealers here and seeing like
amazing, amazing results.
Kyle Mountsier (18:40):
Explain the agentic, AI, just a little bit,
because I think some people getlost with that, like, explain
exactly what it's doing and whatyou're hoping to have it
accomplished. Because when wetalk about AI, a lot of people
are talking about, like, theability to converse with it, but
you're talking about takingaction. So go into like, what's
like in a specific use case,what are you what are you seeing
(19:01):
the agent, agentic. Ai, doingso.
Sanjay Varnwal (19:04):
Agentic. Ai, I mean conceptual level, it is
like you do set of things andthat that lead to certain
results happening, right? AgentA is just automating all of
these things, where, instead ofa human, that particular agent
is doing things and thenpresenting you with the results
that, hey, I have done this.
These are the results. I mean,if you want to set auto rule
(19:27):
that, okay, if these results areabove 90% confidence, just go
ahead execute them. Or I needto, like, check with it and do
the action that, whether I needto take it or or reject it,
right? So, Agent decay ingeneral, is automating lot of
stuff that you do. Instead ofyou working on the software, the
software is working for you, andit is telling you the options
(19:50):
that, okay, I've done this. Doyou want to present it? Do you
want to publish it? Even thatcard could be automated, right?
So, merchandising, we are a.
Doing a lot of agent again, likewe did the shot, instead of
users signing off, we are, wehave we are. We are basically
signing off all of these things.
We are changing the cars imagesin real time on website, without
informing the dealer, right,based on which image is doing
(20:13):
better. So some of these, thesethings, we are, we are
automating on its own, becausewe are controlling the website
infrastructure now for media andsimilarly here. So some of these
things, right?
Kyle Mountsier (20:25):
So like, which, which image should be the first
image, or something like that,
Sanjay Varnwal (20:29):
one of the first use cases, right? When you get
into pricing, you will be ableto when you manage the ads.
Let's say there is an agent, awhich is, which is AD, which is
contextualized or trained on theads behavior, so it
automatically will will increaseor decrease the budgets or
choose the mediums where itneeds to spend more based on how
(20:50):
the clicks are performing,right? So you don't need a
person to do all these thingsfor you. The agent Aki, which is
trained on those use cases, willdo it for you, and it is way
more efficient.
Kyle Mountsier (21:01):
Yeah, I think, I think that the name of the game
with AI is efficiency, and it's,how do I take the tasks that can
be repeated, understood when itcomes to data, and hand it off
to something that can do itfaster and more efficiently, to
free my people up to do thething that they're best at?
Sanjay, thank you so much forspending time with us today. We
(21:22):
We walked the gamut from like,what you saw on Amazon and what
we saw in the early 2000s tonow, what what you're doing and
what you're what you'reinnovating on top of and I can't
wait to see what you and spinedo as you deeper dive into
automotive thanks for joining ustoday on Auto Collabs.
Sanjay Varnwal (21:40):
Yeah, thanks a lot for having me here, Kyle,
Paul and Mike. Great having chatwith you. You guys are like huge
father.
Paul J Daly (21:52):
One of the things I'm super impressed with is the
hyper focus of Sanjay and histeam to fix small problems, and
how all of those things add upto fixing big problems.
Literally, that sounds maddeningto me, focusing that long on the
problem of reflections, but it'sindicative of the fact that they
understand that the small thingsimpact the big things. And we
(22:14):
see that a lot, even as wedeploy AI, quite a bit within
our business. If you don't focuson the small things, the big
things end up being terrible.
Michael Cirillo (22:22):
This is, like, the worst part of our human
nature, which is that we want toperpetually be at the top step
already. Yeah, we never want toclimb the steps to get there,
not realizing that every stepgives us a lesson needed to
actually stand at the top step.
But like, you want a real world,very painful implication of why
people need to focus on thesmaller thing and be great at
it. Look at that dude. I justwatched that documentary about
(22:45):
the guy the deep blue ocean,whatever
Paul J Daly (22:50):
the oh, he went down and like that. The whole
warnings
Michael Cirillo (22:54):
of like, you need to focus on that thing and
make that thing better. And nextthing you know, it's like,
implodes in the middle of theocean, right? And, and we see
that, and we're like, what anidiot. But we're doing these
things in our businesses everyday. We're ignoring, you know,
(23:14):
the small thing. And so to yourpoint about that's very
inspiring to have thediscipline. And
Kyle Mountsier (23:19):
I'm asking chat GPT, how do I change the world
today, right? It's right. How doI make sure that this employee
is able to deploy this specifictask and function well within my
organization, those type ofspecific questions like, How do
I make sure that mirrors show upevery time I do a background
(23:42):
replace. You know, that's howspecific we have to be in
everything, and especially whenyou're deploying AI tools across
your organization. Doesn'tmatter if you're like doing
communicative, communicative AI,agentic AI, you know, photo
backgrounding, whatever it maybe, being very specific with the
task that that person ortechnology is accomplishing. Is
(24:05):
it going to be extremelyimportant as we move forward,
Paul J Daly (24:08):
fewer better, there you go. Well, thank you to our
friend Sanjay and the team atspine for putting their heart
soul and all the intention tomaking this industry better. On
behalf of Kyle Mountsier,Michael Cirillo. Michael
Cirillo, new Ray Ban, metaglasses and myself. Thank you so
much for joining us on AutoCollabs.
Unknown (24:27):
Sign up for our free and fun to read daily email for
a free shot of relevant news andautomotive retail media and pop
culture. You can get itnow@asotu.com that's asotu.com
if you love this podcast, pleaseleave us a review and share it
with a friend. Thanks again forlistening. We'll see you next
time.
Paul J Daly (24:58):
Welcome to autocala. US,
Unknown (25:03):
recording you.
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.
The Joe Rogan Experience
The official podcast of comedian Joe Rogan.
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Special Summer Offer: Exclusively on Apple Podcasts, try our Dateline Premium subscription completely free for one month! With Dateline Premium, you get every episode ad-free plus exclusive bonus content.