Latest / Code Story | Startup Podcast for CTOs, CEOs and Technical Founders / S13 Bonus: From Airbnb's Chronon Engine to Enterprise Real-Time Feature Compute with Varant Zanoyan, Co-Founder & CEO of Zipline AI

Transcript

S13 Bonus: From Airbnb's Chronon Engine to Enterprise Real-Time Feature Compute with Varant Zanoyan, Co-Founder & CEO of Zipline AI · Code Story | Startup Podcast for CTOs, CEOs and Technical Founders

• 0:01
  [SPEAKER_03]: It was built over the course, I want to say a little over and corner, not quite a house, like maybe four or five months.
• 0:08
  [SPEAKER_03]: I then, if you look at the tools I've used to build this, a lot of the challenge here is just in data computation, data processing, right?
• 0:14
  [SPEAKER_03]: So like you got to take raw data and turn it into useful signals, like features and embeddings, and then use that to backfill.
• 0:20
  [SPEAKER_03]: training data and use that to power online inference.
• 0:22
  [SPEAKER_03]: So it's like those are the best kind of a skill with things.
• 0:24
  [SPEAKER_03]: You take data, you transform it, and you need to serve it with low latency and high freshness in the online environment to power a model inference in real time.
• 0:32
  [SPEAKER_03]: And then you need to do these big batches of backfills that can be slow and expensive if you don't engineer things correctly, but ideally they're fast and cheap.
• 0:39
  [SPEAKER_03]: My name is Varan Sanoyan, and I am the co-founder and CEO at SupplyAI.
• 0:49
  [SPEAKER_02]: a podcast bringing you interviews with tech visionaries.
• 0:52
  [SPEAKER_02]: Six, six months moonlighting goes.
• 0:53
  [SPEAKER_02]: It's the last and all of the backgrounds who share what it takes to change an industry.
• 0:58
  [SPEAKER_00]: I don't exactly know what to do.
• 1:00
  [SPEAKER_02]: It doesn't go as to get right.
• 1:02
  [SPEAKER_01]: who built the teams that have their bad company is its team's help each other, which is proud of our team.
• 1:09
  [SPEAKER_02]: Keeping scalability top of mind, all that infrastructure was up there.
• 1:13
  [SPEAKER_02]: Yes, we've been fighting it as we grew up.
• 1:14
  [SPEAKER_02]: Total waste of time.
• 1:15
  [SPEAKER_02]: The stories you don't read in the headlines.
• 1:18
  [SPEAKER_03]: It's not an easy thing to achieve.
• 1:19
  [SPEAKER_02]: To get yourself a deficit of off, try to begin to ride the ups and downs of the start-up line.
• 1:26
  [SPEAKER_03]: Need to really want it.
• 1:27
  [SPEAKER_02]: Not just about technology.
• 1:29
  [SPEAKER_02]: All this and more on code story.
• 1:32
  [SPEAKER_02]: I'm your host, Newell Apphart, and today, however, aren't suddenly as built a secure way to power your AI with fresh and fast data deployed on your cloud.
• 1:49
  [SPEAKER_02]: Varant Zenoion was born in Washington, D.C. and grew up there and in Switzerland as well.
• 1:54
  [SPEAKER_02]: He now lives in the Bay Area, specifically San Mateo.
• 1:57
  [SPEAKER_02]: He spent time at Palantir Technologies as well as a Stint Building ML Tech at Airbnb.
• 2:03
  [SPEAKER_02]: But outside of tech, he loves the outdoors, tending to his garden of plants and vegetables.
• 2:08
  [SPEAKER_02]: After taking good hike, he's been digging a good uni pizza for dinner.
• 2:16
  [SPEAKER_02]: Prior to his current venture, Vermont was working at Airbnb, developing Cronin, an open-source data management engine, used to power AIML infrastructure.
• 2:26
  [SPEAKER_02]: It was then that he and his team realized that building and managing data pipelines was a bottleneck for AI dev and decided to spend into a standalone platform.
• 2:38
  [SPEAKER_02]: This is the creation story of Zipline AI.
• 2:46
  [SPEAKER_03]: So Zipline AI is a feature and embedding platform for machine learning.
• 2:50
  [SPEAKER_03]: So what that means is that data scientists and engineers who are building machine learning models use of my AI to build those models better to automate a lot of the complex, difficult infrastructure challenges that are required to power those models in production and to basically use more data improve model performance and improve the speed at which they're able to iterate on those models.
• 3:12
  [SPEAKER_03]: The story has to tell me how we got here.
• 3:14
  [SPEAKER_03]: It's hard at Airbnb, where I was working with machine learning teams.
• 3:19
  [SPEAKER_03]: So the trust team was one of the first ones that we worked with, and they did all sorts of fraud detection ML.
• 3:24
  [SPEAKER_03]: So there was a lot of payment shapes, fraud challenges at the time.
• 3:29
  [SPEAKER_03]: Every reason for Alabama in many ways is also a payments company.
• 3:32
  [SPEAKER_03]: It moves money across bank account, across borders.
• 3:35
  [SPEAKER_03]: And as a result, it's the target of a lot of the same for fraud attacks on the traditional and processors.
• 3:42
  [SPEAKER_03]: people trying to extract money from the system fraudulently in any number of plays, and they want to take it right faster.
• 3:49
  [SPEAKER_03]: They want to use more real-time data, dreaming features, capture signals, of changing user behavior and real-time and use that to protect these bad actors, better, and all those things were really difficult.
• 3:59
  [SPEAKER_03]: Basically, that was the first these chase that.
• 4:01
  [SPEAKER_02]: We supported it, and so I was taking a look at that and seeing, like, okay, why is this so hard or are the ball next in?
• 4:06
  [SPEAKER_03]: We identify that essentially the data infrastructure, the data pipelines, our use, these models, both for training and serving, is basically where the majority of the complexity and the friction was going in.
• 4:17
  [SPEAKER_03]: And so we started building tools, I would turn on the see if we could take that iteration loop for months, some better, and we eventually got it down for months of days.
• 4:24
  [SPEAKER_03]: That was a pretty big breakthrough.
• 4:25
  [SPEAKER_03]: And
• 4:27
  [SPEAKER_03]: After that, there was other machine learning use cases at Airbnb, like the ones that, well, my income actually first, we think of just like search and personalization and search ranking.
• 4:35
  [SPEAKER_03]: And so that became the second video space that we worked with and that was challenging for some new reasons, meaning that scale is even greater, the latency requirements are even tighter.
• 4:44
  [SPEAKER_03]: And sometimes you're ranking thousands of listings, and every time someone clicks that drag on Airbnb.com, on the app and drag said, that's a new search that happens in a new geobox.
• 4:52
  [SPEAKER_03]: There's new listings that need to be ranked and we want those to be personalized for each user that's searching so that they can show you things that are good for you.
• 4:59
  [SPEAKER_03]: That kind of took the stale of the challenge up, but at the end of the day there's a similar thing, right?
• 5:03
  [SPEAKER_03]: So you want to use other models, have a goal.
• 5:06
  [SPEAKER_03]: I want to be able to loop and you do all that goal.
• 5:08
  [SPEAKER_03]: It's either trading dollars, yeah, adding fraud or easing dollars.
• 5:11
  [SPEAKER_03]: Then you would do some more work over the next like year or so to support the search team better and eventually they were happy.
• 5:16
  [SPEAKER_03]: And then this novel is rolling and this infrastructure within Airbnb that would later become known as Kronon, which is the open source when we eventually opened the source state.
• 5:24
  [SPEAKER_03]: It was getting more adoption and then we started talking to the engineer and team that striped the payment costs of company.
• 5:29
  [SPEAKER_03]: And they were struggling with similar challenges and that's one of the idea this open source came about.
• 5:34
  [SPEAKER_03]: And so we collaborated with Stripe across company bounds and opening for even further.
• 5:38
  [SPEAKER_03]: I don't know when fully open source and then Netflix and OpenAI and a bunch of other big companies started using it.
• 5:44
  [SPEAKER_03]: So it was the open source or some very healthy state.
• 5:47
  [SPEAKER_03]: Yeah, that's why we eventually started the company around around this core challenge.
• 5:51
  [SPEAKER_03]: The vision goes remain the same.
• 5:54
  [SPEAKER_02]: Let's move forward into the MVP then.
• 5:57
  [SPEAKER_02]: So that first version that you built, talent would take to build and what sort of tools we're using to bring it to life.
• 6:04
  [SPEAKER_03]: The first version, this was Airbnb circa 2017, 2018, was, it was built over the course of, I want to say a little over a corner, not quite a half, like maybe four, five months.
• 6:19
  [SPEAKER_03]: I then, if you look at the, the tools I've used to build this.
• 6:22
  [SPEAKER_03]: A lot of the challenge here is just in data computation data processing, right?
• 6:25
  [SPEAKER_03]: So like you got to take raw data and turn it into useful signals, like features and embeddings, and then use that to backfill training data and use that to power online inference.
• 6:33
  [SPEAKER_03]: So it's like those are the best kind of a scope of things.
• 6:35
  [SPEAKER_03]: You take data, you transform it, and you need to serve it with like little latency and high freshness in the online environment to power model inference in real time, and then you need to do these big batches of backfill that can be slow and expensive if you don't engineer things directly.
• 6:48
  [SPEAKER_03]: ideally they're fast and cheap.
• 6:49
  [SPEAKER_03]: When I think of like the main pieces of interest pressure that we used to build this platform, there's the offline batch competition piece and that was initially in the first version was written in high-spart and then there's the online piece which is which the first version was written in spark streaming and now we're no longer on high-spart we're on just if it's been a less fart and a online pat for you are on flank for data processor.
• 7:12
  [SPEAKER_03]: So there's been a bit of an evolution there where a lot of those pieces were, eventually re-evaluated in the results.
• 7:17
  [SPEAKER_03]: But the one piece that hasn't changed is the user-facing API.
• 7:20
  [SPEAKER_03]: That was always in Python and still isn't Python.
• 7:22
  [SPEAKER_03]: That's kind of the language where data science has lived.
• 7:25
  [SPEAKER_03]: And that's for a lot of that.
• 7:26
  [SPEAKER_03]: Embellierity is, so you need me users for the red on that side.
• 7:30
  [SPEAKER_02]: Let's stay on it for a minute, what I'd like to kind of dive into is a core decision you had to make and building that.
• 7:35
  [SPEAKER_02]: That's that could be around technology selection, could be around approach, could be around something that maybe is not obvious from the outside, but I'm curious about one of those core decisions you had to make and how you cope with it.
• 7:46
  [SPEAKER_03]: I think the biggest one was right at the start.
• 7:49
  [SPEAKER_03]: And I think this is one thing that differentiates us from other pieces of infrastructure in this space, which is that we are going to tackle the competition problem.
• 7:58
  [SPEAKER_03]: If you look at the features of our space, there's kind of different architectures.
• 8:02
  [SPEAKER_03]: One that I can one is like, like, the call, bring your own compute, where it's like, hey, you handle computing your features.
• 8:06
  [SPEAKER_03]: This is a AV store, plus maybe some metadata around it.
• 8:09
  [SPEAKER_03]: We decided that that was not the right.
• 8:11
  [SPEAKER_03]: And we decided that based off of really looking at the use cases that we were supporting at Airbnb,
• 8:22
  [SPEAKER_03]: The downside is that it's a much more difficult skill to tap, right?
• 8:26
  [SPEAKER_03]: And saying, okay, this is the usability you guys handle your features up to you, the rest of it.
• 8:31
  [SPEAKER_03]: It's tempting because it's far simpler, but it really doesn't solve the core of the user challenge.
• 8:36
  [SPEAKER_03]: I think we saw that because we were very close to users.
• 8:39
  [SPEAKER_03]: And we were perfectly worth building a platform that we wanted to have as we were so lucky.
• 8:43
  [SPEAKER_03]: These various teams, these are their community.
• 8:45
  [SPEAKER_03]: And then for users a whole side of a headache from an infrastructure, really.
• 8:48
  [SPEAKER_03]: But the nice thing is that I had an headache that you're taking on yourself,
• 8:52
  [UNKNOWN]: You
• 8:53
  [SPEAKER_03]: But the reason why it's fired is because if you look at the way data and pressure as it falls, I'm sorry with things like first there was high, and before we had spark yet high, even at Airbnb when I was there, and then there was spark, and it's no flag.
• 9:04
  [SPEAKER_03]: And then there were things like Kafka, message buses, or you did it in just reading data from, and all these things were very separate.
• 9:11
  [SPEAKER_03]: They spend, like the workflow that users had to take stands, these various tools.
• 9:15
  [SPEAKER_03]: And that's a lot of work that they've made from.
• 9:16
  [SPEAKER_03]: Nobody had an end.
• 9:18
  [SPEAKER_03]: There was no end to end, sort of platform.
• 9:23
  [SPEAKER_03]: and so that's what we saw as the gut, that's what we saw as the right skill for the product.
• 9:27
  [SPEAKER_03]: And now is one decision that you made earlier on.
• 9:31
  [SPEAKER_02]: Okay, let's move forward then.
• 9:32
  [SPEAKER_02]: So you've got the MVP done.
• 9:33
  [SPEAKER_02]: You've made the hard decisions.
• 9:35
  [SPEAKER_02]: You're getting some traction and you're validating your assumptions.
• 9:38
  [SPEAKER_02]: How did you progress and mature it from that point?
• 9:41
  [SPEAKER_02]: And I think that it's worth it in a box of what I'm curious about.
• 9:43
  [SPEAKER_02]: How you build your roadmap?
• 9:44
  [SPEAKER_02]: How you go about deciding that okay?
• 9:46
  [SPEAKER_02]: This is the next most important thing to build or to address with Zipline.
• 9:51
  [SPEAKER_03]: The company is two years old, right?
• 9:52
  [SPEAKER_03]: But like I said, the project has been underway for five, six years, right?
• 9:56
  [SPEAKER_03]: And so, first it was all internal, right?
• 9:59
  [SPEAKER_03]: Internal, at Airbnb, building it.
• 10:01
  [SPEAKER_03]: And it's really nice to be able to do those inside of one company and to get a few things wrong.
• 10:05
  [SPEAKER_03]: And so, even just to architecture, a lot of our batch processes and jobs, like how we do that, that algorithm's behind.
• 10:13
  [SPEAKER_03]: And these point in time-to-act features over large backfills.
• 10:16
  [SPEAKER_03]: There's a lot of decisions that are made there that we have to check the review and to those decisions result in an essentially having to completely change sometimes the way out of the data that we store and do these big data migrations and things.
• 10:27
  [SPEAKER_03]: And those were in a difficult things to do inside one company, but they would have been nearly impossible to do if we were doing this whole thing as a plan.
• 10:34
  [SPEAKER_03]: So, it was really nice to be able to add an area of being able to be incubating ground for this thing to have real use cases at scale, and so to make mistakes.
• 10:43
  [SPEAKER_03]: That wasn't simply there by the way.
• 10:44
  [SPEAKER_03]: Every time we made a big change like that, we had to think about how to migrate these teams that were relying on us still.
• 10:49
  [SPEAKER_03]: It's far easier to do it there than I just did do it.
• 10:51
  [SPEAKER_02]: Now, wherever you have all four customers tend to make a big change like that, would be close to impossible.
• 10:56
  [SPEAKER_03]: Early on, the roadmap was driven by, okay, once on fire, once forging what's not, and a lot of that came down, so, okay, how do we scale this topic?
• 11:02
  [SPEAKER_03]: You should mention that we built.
• 11:04
  [SPEAKER_03]: Now that we're a company, the roadmap is very different.
• 11:07
  [SPEAKER_03]: We're the core engine, which is what's open source, has grown on project.
• 11:11
  [SPEAKER_03]: The AI, it's been super battle tested.
• 11:13
  [SPEAKER_03]: It's the only roofed valve.
• 11:15
  [SPEAKER_03]: And now, a lot of our roadmap is completed by, what does the enterprise platform that we're building around this open source engine need to look like to work really seamlessly and really well.
• 11:24
  [SPEAKER_03]: at more companies, and so that looks a lot hard, like integrations, and our private features around access controls, add-ups, or ability, and governance, and things that you need to be notified immediately with companies existing infrastructure, and with their security standards.
• 11:41
  [SPEAKER_03]: So it's a very different category now than it was in the early days.
• 11:47
  [SPEAKER_02]: So how did you build your team?
• 11:49
  [SPEAKER_02]: What did you look for in those people to help you decide that
• 11:57
  [SPEAKER_03]: We did pull in a lot of people that we had worked with both on Airbnb and from Stripe, what we did do.
• 12:02
  [SPEAKER_03]: And for collaboration, we had the chance to work with a lot of really great people and data infrastructure.
• 12:07
  [SPEAKER_03]: Some people that were two experts in the technologies that we use are badges for processing and really deeply understood the space and people that we've worked with personally and that we like the way they work.
• 12:18
  [SPEAKER_03]: Basically, but founding team was from a pretty tight effort that folks have like all worked together already, which is the best thing.
• 12:25
  [SPEAKER_03]: I'd say, you know, but they are, you know, how they work.
• 12:27
  [SPEAKER_03]: That forms a core team, but when you have a core as you like that, I think they're kind of a center of mass and a gravity that people's been working people.
• 12:34
  [SPEAKER_03]: And yeah, I don't know, I don't know how you can't obviously grow that way, you do 1,000 people, but we're at about 10 people out.
• 12:40
  [SPEAKER_03]: So if you want to, if you want to go 10x bigger, one of the same quality that will probably be a new challenge, but we're not there yet.
• 12:49
  [SPEAKER_02]: Let's move into scalability.
• 12:51
  [SPEAKER_02]: This will be interesting.
• 12:51
  [SPEAKER_02]: This is baked into your bread and butter in the way you think.
• 12:54
  [SPEAKER_02]: But I'm curious about how you approach scale from the beginning in the early days.
• 12:59
  [SPEAKER_02]: And then I'm also interested in stories where you've had to fight scale as you've grown.
• 13:04
  [SPEAKER_03]: You know, when we talk about sale, we usually just talk about data stale for the use case.
• 13:08
  [SPEAKER_03]: So I think like I alluded to earlier, the search with the highest scale.
• 13:13
  [SPEAKER_03]: Use case meaning that there is like the most advanced fire and it was data that needs to be trained
• 13:20
  [SPEAKER_03]: In this data processing engine world, there's like, I think one big challenge in terms of scales often skew.
• 13:26
  [SPEAKER_03]: Well, what's do you need to design some keys that you're aggregating by, have more data.
• 13:30
  [SPEAKER_03]: And others, at least, like, you know, a fraudulent high pH rest that's setting up a bunch of traffic to your science.
• 13:35
  [SPEAKER_03]: So we try to compute features for high pH rest.
• 13:37
  [SPEAKER_03]: That one IP address becomes a bottleneck for computation.
• 13:41
  [SPEAKER_03]: Basically, if you don't architect things carefully, then as you're doing this distribution computation, which is basically taking data and sending it across a bunch of machines to ban out a copy to get it done faster, for parallelism, you can end up in a world where a lot of data gets sent to small number of executors or a small number of machine progression machines for processing.
• 13:57
  [SPEAKER_03]: And so it doesn't really matter how far you're planning it out.
• 13:59
  [SPEAKER_03]: You still have these copies that are effectively a copy bottlenecks in your systems.
• 14:04
  [SPEAKER_03]: So, one of the early things that we did when we were supporting the fraud team was architected in a such a way that we could handle skewed data, which was great.
• 14:11
  [SPEAKER_03]: But the funny thing is actually then later on, and this was an advantage that we made in the last couple of years, because we hadn't got that skewed assumption baked in from the start.
• 14:19
  [SPEAKER_03]: We actually realized that a for uns skewed cases, we can actually get much better performance by having a different computation time, a different work.
• 14:26
  [SPEAKER_03]: And so that was like a further optimization that we recently which was called the non-sumode, or un-sumode in the communication engine.
• 14:33
  [SPEAKER_03]: So now we forked how we compute data in that.
• 14:35
  [SPEAKER_03]: I still wonder whether it's you or not a mother we need to do that additional work.
• 14:39
  [SPEAKER_03]: Anyway, this is just like a little peeking as you kind of what that process has within is like, a lot of okay, here's this these days it's not working, let's fix it this way, but then over time, iterating and further fine-tweeting further outweights.
• 14:49
  [SPEAKER_03]: The engine, and this is why it's great to have open source for it because that's a lot of work or one key to do, but when you have Netflix doing it in open AI doing it, I mean, it's applied doing it in an Airbnb doing it in a script doing it, you have all these people that are contributing to this engine making it better than it's asked or making it for it in her heart.
• 15:04
  [SPEAKER_03]: Case is in more data, profile is done.
• 15:06
  [SPEAKER_03]: Everybody is also editing.
• 15:08
  [SPEAKER_03]: Obviously, like this also, it's another reason why I was nice to start this thing off at Airbnb and be in the right that we could have against our endpoint, and then it was nice to go open source so that more and more people could contribute to it.
• 15:19
  [SPEAKER_02]: Okay, as you step out on the balcony, and you look across all that you've built as far with the Zipline AI.
• 15:24
  [SPEAKER_02]: What are you most proud of?
• 15:26
  [SPEAKER_03]: I would just say the feedback that we've gone from users, adding people, so they're producing it.
• 15:31
  [SPEAKER_03]: We were using here, it was really tough.
• 15:33
  [SPEAKER_03]: And now I have, I used to try to avoid using it.
• 15:36
  [SPEAKER_03]: And now what you guys, it's like, I'm actually excited to use it.
• 15:40
  [SPEAKER_03]: And so I could get my heart fixed on the way I have, but we've been visioned done, and it's hearing stuff like that.
• 15:45
  [SPEAKER_03]: It's great to hear it internally from users at Airbnb, but to hear it from customers, that's the one, and it's like very exact.
• 15:53
  [SPEAKER_02]: Let's flip the script a little bit, tell me about a mistake you made and how you and your team responded to it.
• 15:59
  [SPEAKER_03]: Early on, we made a mistake that I could just comment for startups to make, which is to overindex on the first, maybe sometimes it's the first company or the first set of companies that are talking to you, as potential customers, and to drop everything and do what they are kind of indicating, whether or not I need to make it fundamentally like sense, that's the eyes for everything that you could be building.
• 16:21
  [SPEAKER_03]: And it's something, right?
• 16:21
  [SPEAKER_03]: Because you're just getting going, you don't have any customers yet.
• 16:24
  [SPEAKER_03]: And they're asking for this thing.
• 16:26
  [SPEAKER_03]: It's tough because it's obvious why a lot of teams need that to see.
• 16:29
  [SPEAKER_03]: But yeah, we certainly may have done a stay right now, I think.
• 16:32
  [SPEAKER_03]: But we started, you know, we built a part of the staff that honestly saw that to have, and now it's now we have it, and it does add value.
• 16:39
  [SPEAKER_03]: We invested far too much into it, far early on, where we end up other things.
• 16:42
  [SPEAKER_03]: It would have unlocked way more growth for us early on.
• 16:46
  [SPEAKER_03]: But we saw it, we sort of started off down at an invested 4.5 months into going deep on this one area.
• 16:55
  [SPEAKER_03]: We learned from it, but when we realized what was happening and realized how we were getting guided rounds and not the right direction, and it's hard to know when you first start off on it, because you're like, but they will invest a month in this, but then one of the terms to turn into three, and then we'll monitor heads up for air.
• 17:09
  [SPEAKER_03]: And we had to figure out what are the things that are going to generalize to the most.
• 17:13
  [SPEAKER_03]: Users are non-blocked the most early customers to actually get any value in action, action.
• 17:18
  [SPEAKER_03]: And at that point, you also have more people who you're talking to and you have a better view on that.
• 17:21
  [SPEAKER_03]: It's tough to say no, but you learn to say no, a little bit to things that people are asking for early on and the figure roadmap as opposed to let it be dictated by customers who are sometimes much larger enterprises than you are.
• 17:37
  [SPEAKER_02]: Let's look into the future.
• 17:38
  [SPEAKER_02]: What does the future look like for Zipline AI for the product, for where the industry is going, all the things with MLAI and the changes every day, what does the future look
• 17:49
  [SPEAKER_03]: AI, ML is changing rapidly.
• 17:51
  [SPEAKER_03]: I think the states that we're in, I've always focused on these these cases.
• 17:55
  [SPEAKER_03]: It's like risk-positioning, more broadly, which includes payment fraud and includes underwrite-in-discisioning, by now, pay later type use cases, and then there's personalization, include search ranking, Python ranking, Python first-length Asian, let's say first-size issues.
• 18:09
  [SPEAKER_03]: Those two use stations at the end of the day, certain things are not going to change about them, which is a requirements.
• 18:14
  [SPEAKER_03]: They are high QPS, they are low latency, and that is the way they work.
• 18:19
  [SPEAKER_03]: And so as a result, they are their profits or language models.
• 18:23
  [SPEAKER_03]: Basically, where
• 18:25
  [SPEAKER_03]: The industry has changed and there is twofold.
• 18:27
  [SPEAKER_03]: One is more broadly gendered in architecture, so things like gendered recommendations have been coming around there.
• 18:33
  [SPEAKER_03]: So not language allows for things like somatic IDs, which are really interesting, and that's where a lot of our mining value.
• 18:40
  [SPEAKER_03]: And their models in the performance of their systems.
• 18:42
  [SPEAKER_03]: And so we are our infrastructure is built to support with these cases and to meet the notes in the point built with a production easier.
• 18:49
  [SPEAKER_03]: But then, yeah, more broadly, it's like the other thing that we've seen is the same way that AI, and LLMs have accelerated voting.
• 18:55
  [SPEAKER_03]: They're also accelerating building an iterating on these models.
• 18:58
  [SPEAKER_03]: So, it's not that the language models aren't replacing cloud detection models or search models, but they can be hugely impactful for helping people iterate on these models to run experiments to engineer better models and then to get better pipelines.
• 19:10
  [SPEAKER_03]: And so our infrastructure is also built to be very coding agent first, really, that right now, you don't need much, you can use cloud code to write code and to develop features on at applications.
• 19:20
  [SPEAKER_03]: I can do this kind of ML and hear if ML engineering about coding agent needs is set up tools to work with data and view build those data by clients and to view model training and essentially run experiments that will reproduce in production.
• 19:32
  [SPEAKER_03]: When you want to do it, you want it to have an API.
• 19:34
  [SPEAKER_03]: It can run in Fierman, so you can tell us, hey, you run this sort of, it's a huge, we can guide it.
• 19:39
  [SPEAKER_03]: Along with what turns out things you want to try, or what things you want to try, and the security agent had the right tools.
• 19:44
  [SPEAKER_03]: It can be a very powerful assistant for people.
• 19:46
  [SPEAKER_03]: A to rate on these models, so make them better, and to improve the performance of these personalization and current detection systems.
• 19:54
  [SPEAKER_03]: And so that's where our infrastructure fits into this feature, basically.
• 19:59
  [SPEAKER_02]: Let's switch to you.
• 20:00
  [SPEAKER_02]: Who influences the way that you work?
• 20:02
  [SPEAKER_02]: Am I person or many persons or something?
• 20:04
  [SPEAKER_02]: You look up to him.
• 20:05
  [SPEAKER_02]: Why?
• 20:06
  [SPEAKER_03]: The funny thing here is that there's no one really famous.
• 20:09
  [SPEAKER_03]: Like when I think about the people that influence me and the way I work, I don't think of what are the celebrity people and technology.
• 20:17
  [SPEAKER_03]: It's warm just, oh, so I worked with an Airbnb and thought of how much you're before that.
• 20:22
  [SPEAKER_03]: And there's a handful of people there and they're not even necessarily the highest runs of the corporate ladder or anyone that the public would have heard of, but just people who I part with.
• 20:30
  [SPEAKER_03]: And when I think about why I only come to them and why I learned from them.
• 20:34
  [SPEAKER_03]: And sometimes they, yeah, like I said, sometimes they were not even like they were not necessarily the CEOs or whatever.
• 20:38
  [SPEAKER_03]: Sometimes they were literally my peers in terms of level.
• 20:41
  [SPEAKER_03]: But it still ultimately didn't the right thing.
• 20:44
  [SPEAKER_03]: They didn't get caught up in corporate politics kept eye on what the right students should be thought about once right
• 20:52
  [SPEAKER_03]: And kept it simple and what I mean by that is it's easy to lose sight of what matters But it's if you really prioritize what's going to be valuable for the company What's the valuable for the user and do the simplest thing the land and value and not to get caught out dead The politics and the scope for end of this and that that and that's something to get bigger is really easy to get caught out of those things
• 21:12
  [SPEAKER_03]: So folks who are able to do that, to take a principal approach to their work, I always admired those folks a lot.
• 21:20
  [SPEAKER_03]: And yeah, as I worked that tried to, at least focus on that.
• 21:24
  [SPEAKER_03]: And it's easy.
• 21:25
  [SPEAKER_03]: I think it translates a bit to, as you're running your own company, it's now you don't have, sure you don't have core politics to play anymore.
• 21:31
  [SPEAKER_03]: But there are always things that you could optimize for that.
• 21:33
  [SPEAKER_03]: I'm not the right thing.
• 21:34
  [SPEAKER_03]: Even things that are important for the business, like fundraising or whatever.
• 21:37
  [SPEAKER_03]: It's like you still have things to prioritize, and it's like what do you choose to prioritize?
• 21:42
  [SPEAKER_03]: And I think that we built valuable technology by prioritizing the right things by prioritizing.
• 21:47
  [SPEAKER_03]: Well, it moves the needle on the business, like a little, but will ultimately drive dollars for either via saving money from fraud, buy, line, or if I'm proving bookings and approving search results at the top line.
• 21:58
  [SPEAKER_03]: And by making people more productive, like what tools do people need to build these models better, and to do this better, like, will actually make them, and I can work productive.
• 22:06
  [SPEAKER_03]: It's really easy.
• 22:06
  [SPEAKER_03]: It seems obvious, but it's really easy to be excited those things, as in the chaos of the day to day.
• 22:11
  [SPEAKER_03]: So, if you put in my writing, we're really good at even the eye on the wall.
• 22:16
  [SPEAKER_02]: Okay, last question, so you're getting on a plane and you're sitting next to a young entrepreneur who's built the next big thing.
• 22:21
  [SPEAKER_02]: They're jazzed about it.
• 22:22
  [SPEAKER_02]: They can't wait to show it off to the world and cut my shirt off to you right there on the plane.
• 22:26
  [SPEAKER_02]: What advice to give that person having gone down this road a bit?
• 22:30
  [SPEAKER_03]: I think if there are already at that point where they have something, and they're super excited about it, they're already 90% of weight there, and what I mean by that is to get to that point, there's something in the world, and there's very much water in them.
• 22:42
  [SPEAKER_03]: They're like, okay, we simply can't let this be delayed, it is there needs to be a better way to do this.
• 22:46
  [SPEAKER_03]: There needs to be a better solution for this thing.
• 22:48
  [SPEAKER_03]: Harvard can't let this sort of wrong in the world continue.
• 22:51
  [SPEAKER_03]: And this is my vision for the solution for it.
• 22:54
  [SPEAKER_03]: If they're at that point then I would say, yeah, all they got to do is keep their eye on the ball in that way.
• 22:58
  [SPEAKER_03]: I don't lose sight.
• 22:59
  [SPEAKER_03]: I think that's driving that.
• 23:00
  [SPEAKER_02]: They get to this point and there's clearly something that's driven them to get to that point.
• 23:03
  [SPEAKER_03]: And as long as they stay focused on that, I think they'll all be fine.
• 23:06
  [SPEAKER_03]: If I saw someone like that, I'd be really, really resistant as driving you.
• 23:09
  [SPEAKER_03]: And so just make up every day and do what you think is the right thing to do to solve a problem or make things better for people in that particular way.
• 23:16
  [SPEAKER_03]: And that's literally they're calling.
• 23:18
  [SPEAKER_03]: And so I think they're already there.
• 23:19
  [SPEAKER_03]: I think I wouldn't say all I would say is you're already there.
• 23:21
  [SPEAKER_03]: You're already doing it just on the inside of it.
• 23:23
  [SPEAKER_03]: That's all you gotta do.
• 23:24
  [SPEAKER_02]: That's awesome advice and super encouraging.
• 23:27
  [SPEAKER_02]: Well, Vermont, thank you for being on the show today.
• 23:29
  [SPEAKER_02]: Thank you for telling the creation story of Zipline AI.
• 23:32
  [SPEAKER_03]: Thank you.
• 23:33
  [SPEAKER_03]: Thanks for having me.
• 23:33
  [SPEAKER_03]: It's been fun.
• 23:36
  [SPEAKER_02]: And this concludes another chapter of Code Story.
• 23:50
  [SPEAKER_02]: code story is hosted and produced by Noah Labhart.
• 23:53
  [SPEAKER_02]: Be sure to subscribe on Apple podcast, Spotify, or the podcasting app at your choice.
• 23:58
  [SPEAKER_02]: And when you get a chance, leave us a review, both things help us out tremendously.
• 24:06
  [SPEAKER_02]: And thanks again for listening.