← Digital Social Hour
Digital Social Hour
Alan from Tryolabs - AI Is Not Magic: Here’s What Real Deployment Looks Like | DSH #1768
00:18:38
We’re at the AI4 Conference with Alan from Tryolabs — an AI consulting firm that’s been building real machine learning solutions since 2009, before AI was mainstream. Alan breaks down what “AI in the real world” actually looks like: not flashy demos, but messy enterprise systems, outdated data layers, and the hard work of making models reliable at scale. We also dive into the most inspiring side of the industry: AI for good — using computer vision to map schools via satellite imagery for UNICEF, helping sustainable fishing become measurable in real-time, identifying lions by whisker patterns to protect conservation land, and detecting early wildfire smoke before it becomes a crisis. If you’re building in AI (or thinking about it), this episode is a reality check — and a roadmap. What You’ll Learn 🦁 How AI can identify lions by whisker patterns to support conservation 🛰️ How satellite imagery can help map schools in developing countries 🌡️ How AI can measure risks like heat ...
Loading summary
Transcript80 lines
- [00:00]
A
Like within the species. I can tell like if this lion is Bob or if this lion is Alice based on their whisker pattern. So they will conserve the whisker patterns since they're cubs. And the idea here is like these organizations have a data set of many photographers of lions taken through the years. And whenever there's a new photo, we can actually use AI to match it to the existing database. So they can see, oh, this lion was actually found in another reservoir like hundreds of miles away and now it's moved here. So that gives them data to be able to protect those lands.
- [00:28]
B
So fascinating.
- [00:28]
A
Yeah. And those are kind of the use cases of AI that AI for good of the world. That it's something that is really interesting for us.
- [00:42]
B
Okay, guys, we got Alan from Trial Labs. Here we are at the AI4 conference. Great to meet you, Alan. What is.
- [00:47]
A
Thanks so much.
- [00:48]
B
Yeah, what is Trial Labs about?
- [00:49]
A
So Trial Labs is an AI consulting and services company. We've been in the space for 15 years. So way before AI was a thing that everybody was talking about. In fact, AI was called machine learning back then. So we started around 2009, 2010, initially serving clients in the Bay Area startup San Francisco, which was the only place in the world where there were crazy enough founders to kind of use these technologies. And through the years we've grown. Nowadays we serve mostly corporate and enterprise clients and also some big nonprofit organizations.
- [01:17]
B
And what is the exact service that you're giving these clients?
- [01:19]
A
So we help them build custom solutions using AI and data to get some desired outcome. In general, it's like business outcomes, but it can also be some outcome that's good for the world or some other initiative that they're pursuing. So yeah, these companies have massive amount of data. There's a lot of things that they need solved. And the application of AI for these problems is something that's actually non trivial. And it's not just a technical aspect. There's of course a big technical aspect to that, but there's a lot of processes, things that have been always done in a certain way. There's people involved and there's like complex systems and you need to make AI fit in those complex systems and actually provide business value for them.
- [01:58]
B
Interesting. And you guys have been featured in a lot of major outlets for solving interesting problems.
- [02:02]
A
Yeah, we actually have, depending on specific initiatives that we pursued, there's some that are pretty high profile. Yeah, sorry, there's initiatives that are pretty high profile. So we've done some work with unicef, for example, around the impact of Heat waves on developing children. We've also done some other things that are super interesting. For example, a couple of years back, we did an initiative with UNICEF and we published a paper on this, which was how do you use satellite imagery to detect schools? So the problem that the world has is that in most developing countries, the schools are not 100% mapped by the government.
- [02:41]
B
Wow.
- [02:42]
A
So there's countries in which they don't know more than 50% of the schools where they're at.
- [02:46]
B
That's crazy.
- [02:47]
A
So UNICEF has this initiative called giga, and their intention is to find these schools and go and connect them to the Internet so that these children can have the best education possible. And what we were doing is taking satellite imagery and training custom computer vision models to actually understand what buildings are actually likely to be schools and map them.
- [03:04]
B
That is fascinating.
- [03:05]
A
Yeah. We're also doing some interesting work with an organization called the Nature Conservancy around the sustainable fishing practice. So I think it's like close to 38% of all the fishing stocks in the world are overfished. And that is a big problem for conservation of ecosystems. And there's also, like, push from governments in the regulation of electronic monitoring, and also push for. From some retailers, like Walmart, for example. They announced that every fish that will be in the shelves in the next year or two will have to be fished sustainably. So what we're doing there is like, we're putting cameras on the fishing vessels, and whenever they're fishing, we can classify, like, what's catch, like the intended catch. And they buy catch and get independent metrics.
- [03:45]
B
Wow.
- [03:46]
A
So, yeah, this is an industry that. The electronic monitoring industry exists for a long time. But the problem is that the review cycles for the videos takes like, several months. So you ship a hard drive and then three months later. Yeah, you have overfished here. And clearly you cannot act rapidly with that. So we are shortening the cycles and making sure that this ship can report independent metrics in real time.
- [04:06]
B
Yeah, 38% is a lot. It's a lot. Is that a worldwide issue or.
- [04:09]
A
That's a worldwide issue for sure. I was reading the other day that even in the. That is like putting selective pressure on some species. So the fish are actually shrinking due to the overfishing. So imagine that these smaller individuals, they can kind of escape the nets. So that makes the species overall, like, through the years smaller. And that also, of course, causes problems for conservation and.
- [04:29]
B
Yeah, interesting. So that's a major problem. You guys also track some lions too, right?
- [04:34]
A
We have Done that too. There's an organization called Lion Guardians. And basically the issue is like, conservation is in Africa. They need to understand where the lions roam to be able to protect those lands and to prevent human development in there. And the issue is that tracking lions is a complex endeavor. There's two tracking mechanisms. One is very invasive. You put a collar on the lions. So you have to go there. Yeah, good luck with that. You have to go there, sedate the animal. The colors are expensive. Then you need to replace the battery every year. And there's the other method, which is tracking, that's non invasive. So photographers will go with a telephoto and capture these photos from very far away. And it turns out that the lions, they can be uniquely identified. So this is like within the species. I can tell if this lion is Bob or if this lion is Alice based on their whisker patterns. So they will conserve the whisker patterns since they're cubs. And the idea here is like, these organizations have a data set of many photographers of lions taken through the years. And whenever there's a new photo, we can actually use AI to match it to the existing database. So they can see, oh, this lion was actually found in another reservoir like hundreds of miles away, and now it's moved here. So that gives them data to be able to protect those lands.
- [05:44]
B
So. Fascinating.
- [05:44]
A
Yeah. And those are kind of the use cases of AI, that AI for good of the world. That. It's something that is really interesting for us.
- [05:51]
B
One of my favorite ones was my friend Walter o'. Brien. He used AI to solve the Boston bomber at the marathon. Wow. Searched through thousands of hours of footage and they were able to find his patterns on how he was reacting because everyone else was running away while he was acting casually. Wow. Crazy, right?
- [06:06]
A
Yeah. That's fantastic. I mean, you also hear the kind of dystopian stories about like, what the potential of, of that technology is. Like, look at China, what they're doing.
- [06:14]
B
Mass surveillance, Massive surveillance.
- [06:16]
A
Yeah. But I feel like underlying the, the, the, the underlying technology is like, not bad or, or anything. Like, if it's used for a good purposes, I think it can make a huge difference.
- [06:24]
B
Agreed. Yeah. I am worried about mass surveillance. 1.
- [06:27]
A
I think everybody is.
- [06:28]
B
They got cameras everywhere, all the traffic lights.
- [06:30]
A
They do, they do. I mean, there's, there's of course another aspect of the AI that's called edge AI. So there's like some specific algorithms that can run on device. So for example, retailers can use that for analytics within the store. And that doesn't necessarily have to identify any individual person. They will just count foot traffic, understand, like even classify gender or something that the retailer might care about around how people are moving on their stores. And that does not mean that the fact that there's a camera does not mean that you're being surveilled. This can be something that runs on device and then just like, reports, aggregate statistics. So it depends on how you actually implement that.
- [07:06]
B
Yeah, that, to me, is more, I guess, approachable. Right? Yeah. Rather than someone watching you all day wherever you are.
- [07:12]
A
Correct.
- [07:13]
B
Yeah, I'm good on that. You also used your technology to track some fires.
- [07:17]
A
Yeah, we did, actually. We worked a couple of years back. So everybody knows that in California, like, the wildfires are a massive problem that not only jeopardize, like, property and value, but also human lives. This affects community deeply. And we worked for this startup in San Francisco that went on to raise a lot of money in developing a system that can detect the early signs of wildfire. So these companies installing cameras in many different locations, they can actually triangulate. So where we see signals of smoke, they can triangulate and call the fireman. So in the very first moments of the fire, and if you can actually detect it on the first five to 15 minutes, the chances of it not becoming a massive wildfire is extremely incremental. So the idea here is that with AI, we could train custom computer vision models to detect signals of smoke in the early minutes of the fires that are very difficult to detect for the human eye. So it's interesting that if you see the photos of how they look like, you would never be able to tell there's a fire. But with that, the algorithm can say, oh, there's something here. And when you see the video play out, yeah, there's a tiny signal of smoke over there.
- [08:26]
B
Wow.
- [08:27]
A
And the interesting part is, like, how do you differentiate smoke from fog or from somebody doing a barbecue? Like, there's a lot of challenging aspects that come with developing these models. Like, you start to generate some false positives, and if your system generates too many false positives, then nobody will pay attention to that.
- [08:43]
B
No one will trust you.
- [08:44]
A
So you have to kind of balance it out in a way that it's actually useful, but it can be really accurate. And these companies expanding internationally, they're having great success.
- [08:52]
B
Nice. So the one that happened last week in California, did it track that one?
- [08:56]
A
I'm not sure. We developed the initial version of that solution and then left them to continue, because that's something that we do it's not just about building the solution and staying there forever. It's like we can do what's called project delivery with knowledge transfer. So we will train their own teams to continue developing the solutions through the years.
- [09:12]
B
What was the most difficult model to train out of your 15 years doing this?
- [09:16]
A
That's a great question. I think the fire one is particularly tricky because it's like the data collection effort that needs to undergo of that in order to have a nice ratio of false positive that is manageable is really tricky. There's like any single phenomena, there's a bird on the camera that it can really. Things that you don't think about can happen also, like lens flare and again, fog or some specific pattern of clouds. And then you take your model to another location and maybe there's snow. And when you train in California in these parts, there's never. The models has never seen snow. So there's a lot of custom things that can go wrong and there's a long tail of data that you need to collect for it to be really accurate. So I'd say that's an interesting one. Of course, there's many more areas in which we've worked, so there's work that we've done in forecasting. How can a retailer forecast how many items they need to buy so that they don't run out of stock next week? And there's a lot of complexities in that kind of scenarios. But yeah, I don't know if there's like a single most difficult model, but this is something worth mentioning.
- [10:20]
B
I'm sure you're working on so many different models at the moment, right?
- [10:22]
A
We are, yeah.
- [10:22]
B
I bet. Business is really booming.
- [10:24]
A
Business is thriving. So we're having. I mean, we've been in this space for 15 years and that's every company under the sun right now says they do AI and they're all like mostly users of ChatGPT and similar APIs, like Gen APIs.
- [10:36]
C
Shout out to today's sponsor quote Quince. As the weather cools, I'm swapping in the pieces that actually gets the job done that are warm, durable and built to last. Quince delivers every time with wardrobe staples. They'll carry you through the season. They have fall staples that you'll actually want to wear, like the 100 Mongolian cashmere for just $60. They also got classic fit denim and real leather and wool outerwear that looks sharp and holds up. By partnering directly with ethical factories and top artisans, Quince cuts out the middleman to deliver premium quality at Half the.
- [11:06]
B
Cost of similar brands.
- [11:07]
C
They've really become a go to across the board. You guys know how I love linen and how I've talked about it on previous episodes. I picked up some linen pants and they feel incredible. The quality is definitely noticeable compared to other brands. Layer up this fall with pieces that feel as good as they look. Go to quince.comdsh for free shipping on your order and 365 day returns. They're also available in Canada too.
- [11:43]
A
But we have developed like a very fine understanding of how really machine learning works from the fundamentals and we've done that all over the years. So we have like hundreds of different use cases and yeah, right now we're working for many different industries on the airline business. We're doing models from like optimization of contingency fuel. We're doing revenue management systems, we're doing like gen AI making developers more efficient. We're doing also things in the manufacturing space, in the automotive space. So these companies that have like massive amounts of data, they can.1% for one of these companies is like hundreds of millions of dollars savings every year.
- [12:19]
B
1%?
- [12:20]
A
Yeah, 1%. So it's like a project we did for a major airline last year. It saved them over $120 million and in fuel cost savings. And that is something that, yeah, that project took like a year of development. But that is like the, when you optimize the whole value chain, like massive unlocks happen. And that is like, I think it takes like a change of mindset in the leadership of a company from the top to make sure that there's actually like budgets for R and D and for like what's going on in bigger companies is they have so many opportunities to use AI. However, their data layer is so far behind. And I was telling a colleague that I met right here in the conference that the other day we started a project for a massive retailer. And the way that we got the access to the data set for it was a forecasting project, was a hard drive mailed to our office. So imagine if you want to get the most value out of AI, what can you do if your data layer implies that you have to ship a hard drive rather than giving me cloud access and everything. So there's like, I think we're living kind of in a bubble in the sense that what we see, when we see that AI move so fast, everybody's doing agents, everybody's having the roi, whatever. Like the vast majority of companies are like so far behind of that that they actually need to invest years in this unifying the data layer and make everything ready for extracting the value out of AI.
- [13:43]
B
Got it.
- [13:44]
A
So I think that's the most. The massive part of market is over there.
- [13:48]
B
That's an interesting problem, right?
- [13:49]
A
Yeah, for sure, for sure.
- [13:50]
B
What would your advice be to younger people that are just getting into AI starting companies and getting into this business world?
- [13:55]
A
It's a good question. I think right now it's very easy to be overwhelmed by everything going on. It's very easy to have imposter syndrome in the sense that, yeah, you read the news, you read Reddit, you read X, everybody's doing cool stuff, everybody's getting things shipped, everybody's making a lot of money. And again, it's really easy to say, okay, am I wrong or something? Am I too slow? And I think my answer would be just like, my advice would be try to do things to create a mental model of what the limits of the technology are, where it's best to use it, where it won't actually work at scale. It's very easy to do flashy demos, but then when you go to the real world, there's a lot of complexities. So, yeah, my advice would be don't think you are the imposter. Just, like, go play with the tools, get things done and, yeah, build that mental model of how the future might look like and enjoy the ride. It's a fantastic time to be working in this space.
- [15:00]
B
Great advice. I struggle with imposter syndrome growing up. I think a lot of people in my generation, yeah, it's very common. Yeah, because you're comparing yourself all day when you scroll on social media.
- [15:09]
A
Correct, correct. The grass is always greener on the other side.
- [15:12]
B
That's like, people don't post their bad moments.
- [15:14]
A
Exactly, exactly. I mean, some do, but it's not like, Yeah, I use AI for these 15 different use cases and it actually didn't work for 13 of those who actually posts that. And that is like the reality of everything that's going on behind the scenes. You start understanding, yeah, my demo will be very nicely, but then when I want to scale this, I will hit roadblocks. And I think going back to the enterprise and everybody's promising agents, and there's certainly use cases where agents are really useful, especially if they're kind of narrow. But I think the expectations of some people on the leadership are like, agents are the magic silver bullet that will solve all your problems. And again, an agent that is 99% reliable on a business might be completely useless. Imagine a self Driving car that drives perfectly nicely without crashes 99% of the time, 1 out of 100 will crash. It's useless. So getting to the 99.999999% that you need for the agent to be really self sufficient in a generic scenario is really, really hard. And for many verticals like the technology is not yet there. It might get there eventually, but there's no clear path to that. However, there's maybe like If I use ChatGPT for brainstorming or whatever, I don't care if it just hallucinates some answers to me. I will, I will know, I will read them. Yeah, these are useless. These are useful and I will take that and do my things. But for other use cases that really add business value to the enterprise, you need like a level of real reliability that's not yet there.
- [16:45]
B
Well said. So that being said, do you think a lot of these AI companies are going to fail? Do you feel like we're in an AI bubble right now?
- [16:52]
A
I don't think we are like in a bubble as we were before. I think we are in a period of realization. So they're going to realize that the AI as a silver bullet promise, the magical chatgpt moment does not really translate to I plug in these models and then I don't need to hire more people or then I like we completely revamp my business process. It's like the real world again is super messy and it doesn't run like that. And I think the industries will take time to adapt.
- [17:22]
B
Yeah. You still need that human touch.
- [17:24]
A
Yeah, absolutely. Absolutely.
- [17:25]
B
Probably for years to come. I mean it's hard to predict, but.
- [17:28]
A
It'S very hard to predict. I would bet that for the foreseeable future there's a lot of borrowing timelines for whatever people call AGI. I think that for the next at least like 10 years.
- [17:38]
B
Oh, 10 years. Wow.
- [17:38]
A
Yeah, 10 or maybe even more. Like I think people are the this, this notion about exponential progress might not be such when you are trying to apply all these to a very complex system with a lot of moving parts.
- [17:53]
B
Right.
- [17:54]
A
So even if you had these magical models that can do everything, the all the industries being disrupted might take like a longer time.
- [18:01]
B
You still have to implement it. Yeah, yeah, that's a good point. Wow. And it's been awesome. And where can people find what you're doing and protect work with you?
- [18:08]
A
Yeah, perfect. So you can follow us on LinkedIn, tryolabcry. We also have a website, triolav.com we have a technical blog. We post a lot of interesting content around, like these projects that I've mentioned and others. Like for technical folks, there's like, how do we do these things for business people is how they can help your business. So yeah, I'd be glad if people check it out. And thanks so much for having me.
- [18:29]
B
Absolutely, man. Check them out, guys. We'll link it below. See you next time.
- [18:32]
C
I hope you guys are enjoying the show. Please don't forget to like and subscribe. It helps the show a lot with the algorithm.
- [18:37]
B
Thank you.