A Conversation with Vercel | EuroPython 2026 Platinum Sponsor

[music] >> So, we're here with Yuri from Vercel. So, do you just want to give a quick introduction to yourself, to Vercel, and also what you're doing with the AI cloud? >> Yeah, I'm Yury Selivanov, Python core developer. I've been a core developer since 2013. I worked on many different things in Python. Specifically, I think like the most famous contribution is adding async await syntax to Python and working on async IO, shaping up the async ecosystem. Or being part of that process. I created uvloop and asyncpg. Both are quite popular and widely used. After that, I was building Gel. Some people might know it by name EdgeDB, which was a high-level database built on top of PostgreSQL. I would say Pythonic database built on top of PostgreSQL. And relatively recently, about half a year ago, we were acquired by Vercel. And now we are trying to make Vercel Pythonic, which is an interesting challenge. >> Well, I think acquiring your company is a very good start for them. >> Thank you. >> so, your your CEO has given quite a lot of talks about building the best AI cloud at Vercel. So, what does he mean by that and like what would that look like in practice? >> Essentially, building AI agents in requires a lot of different moving parts. And Vercel fills all the gaps that are there for people when they need to deploy AI application or creating AI application. And the range of solutions that Vercel uh, offering is pretty is pretty wide. Some of them are focusing on more like low-level technical side and some are incredibly high-level. Like, for example, the incredibly high-level solution is V 0, uh, which is a in-browser coding agent. You can open V 0 and you can create a full-blown, uh, application with it relatively quick and easy. And I think it's one of the best ones in the market. It's a huge product on its own. Um, but then we can start talking about like low more low-level aspects of what, uh, what it takes to build an AI application. Like, one of them, for example, is how you communicate to LLMs and different providers. So, AI has, uh, AI, uh, Versel Versel AI Gateway. Uh, and that is basically like a load balancer for, uh, LLM requests. Like, for example, you're using Opus and Anthropic is down, then we're going to reroute you to another provider without your application, uh, experiencing a hitch. It's like it's quite convenient and, uh, the performance is great. And Versel actually makes zero markup on top of that. It's just like it's just a service that that that they give to the community. Uh, so I think it's a no-brainer to use it. And, uh, then a bunch of other things like, uh, AI Gateway is one of the products, but then, uh, when you start implementing agents, you need to start caring about durability. And then you need to run workflows and we have, uh, workflow SDK and the whole, uh, cloud orchestration layer, uh, to, uh, to run those. Uh, AI, uh, Versel has JavaScript AI SDK, which I think is pretty much a de facto standard, uh, for JavaScript. And, uh, my team and I are building Python AI SDK, which is actually quite different from its AI counterpart because it's specifically designed to work with Python. It's designed around Python language. Uh, but on top of that, AI, uh, Versel also has, uh, AI SDK UI, which is a whole ecosystem on its own, which essentially allows you to plug into the AI and LLM events streaming from the server and render in browser which is in browser UI with just like a couple of lines of code. It's the ecosystem is huge and vast. And I can actually continue but like ultimately it all boils down to where you can host your code and how smoothly you can do. And this is like the core Vercel business. We can take your Python code and JavaScript code. We can host it and you can build your agent on top of that and great DX guaranteed. >> [laughter] >> eventually. >> And it's really exciting to hear about companies that are actually getting to the core of the engineering problems because essentially LLMs are extremely massive ML problems and we're seeing the traditional problems we had with that you know, dialed up to a million. And so seeing this sort of maturity, thinking about infrastructure, thinking about load balancing, like these traditional engineering problems as well as accessibility to you know, platform, it's it's it's an exciting time to be working in this space and it sounds like work that Vercel's doing is really like at the forefront of that. >> Yeah, 100% and we we we host so many different companies. I think like half of Fortune 500 probably. So we are exposed to an incredibly wide range of different of different problems that our customers have and some of the biggest companies in the world have. And obviously we're all about closing the gaps and because Vercel is unique like we are not just building the zero, we are not just building web hosting, we are just not just building the the gateway product. We are building all of that at once. We see a lot of unique opportunities how those things should actually be bridged together and be pieced together. How we can build like a coherent DX and DX story around all of this. How we can make sure that like agents understand our infrastructure which is like an important part of of today's offering. So ultimately what it boils down for the user is that they start using Vercel and they can just like ship their idea on Vercel without like spreading their effort between multiple different separate service providers and figuring out how to glue them all together. >> Especially when everything's changing so quickly. Like that really gives you robustness and durability. >> Exactly. Exactly. They basically like any ecosystem needs to be coherent, they need to be composable. And those two things are quite core to Versel. Like how can we ship something that like plays nice with each other? It's like like Lego bricks. And also everything is documented, everything to human like chips and everything like talks to each other properly. >> Even documented. >> Yeah. Well, it's important for Asian sound still humans. >> Absolutely. Like um Yeah, we could go into this whole thing about how documentation is done. >> that just manages documentation and makes documentation look and read the best in the world. Like we are obsessed with this kind of detail at Versel. Like that's >> I love to hear this in 2026. >> It's still very important and will continue to be important. Maybe even more important than code. >> I agree. So, running Python in a serverless environment, it brings a lot of challenges. So, which of these are you most focused on, most interested in, and kind of most concerned with right now? >> I'm interested in all of them because it's like a huge minefield. Like nothing works the way you want it to work. Serverless but Serverless has a lot of challenges with Python specifically because Python is quite heavy, it's a heavy runtime. Python is not famous to be the fastest language. Um and making relatively simple applications work is actually like not that hard. But we have some customers who have like 10 GB of Python dependencies, which is an insane amount. >> Jeez. >> Exactly. Well, it's like huge companies have huge software. And for that just like to start it can take minutes. So, how do you optimize the startup time for that? Like if you want nicely horizontally scalable Python serverless, which basically means that if the load goes up, you immediately can spin up workers in that specific region and uh the and and and the quality of service continues to be high. You basically need to minimize the startup time. And uh that those challenges are pretty fundamental. Like for example, my team uh is um looking very closely at potentially automatically using lazy imports so that we can analyze your code when you when you when you build it and before you deploy it. Yeah. And then we basically have a map which imports in your code base can be made lazy, automatically essentially and safely so. But this is a very hard problem because Python is incredibly dynamic. So you can have the section at a distance where for example, you can have I don't know, like base.py file that defines a metaclass and uh some other completely unrelated code when it when it when it imports, it needs that metaclass. And if that metaclass is not imported, well, it's a the program will just like error out essentially. Uh so it's like building a very complex compiler infrastructure that can that that can find those patterns, understand those patterns, and build a map of like what you actually need to run this specific API handler. Um And then it goes deeper than that because for serverless again, like you need to be able to observe all those things. So we need to build a lot of observability, internal and external observability to understand how how well we are performing. And then again, like for people who have uh hundreds or thousands of dependencies, we need to shorten the build time to give you good developer experience. We need to shorten the deploy time, which basically means that downloading all those packages must be must be as fast as possible and ideally like take zero time. >> Mhm. >> And uh I think we are on the verge of cracking that problem, how we can actually make it so that you're It's irrelevant how how many dependencies you have. Uh but all of those are like pretty challenging engineering problems and uh uh for Python, I don't think like too many companies actually solved them. Um so uh it genuinely feels like just like this super exciting part of engineering world. Uh yeah, so I'm excited about that. >> Actually, have some of the changes like lazy imports 40 and in 314 and 315, have they helped with >> Well, absolutely. Like you know, we're we're building on the shoulders of giants here. Uh they like those changes specifically are more about new applications or existing applications adopting this that patterns explicitly, like starting to specify the lazy keyword. Uh but what we want to do is basically explore if it's possible to do it dynamically so that you don't even have to do it. Because for a lot of programs uh that measure like in millions of lines of code, it it's it's a very tedious process to manually start annotating uh lines of code. It's it's it's hard and tedious even for for agents that can famously sift through millions of lines of code. But again, because like of how intricate these problems are, even for agent without tooling annotating your code automatically which import to be can be lazy and which can can be uh uh like proper runtime bound, it's really really hard. So, um yeah, it's just a hard problem. Uh obviously, we're building on top of the uh of the lazy imports PEP. Um but there is a lot to build still. >> I think I'm starting to see why they wanted to acquire you. >> [laughter] >> Thank you. >> it's always um fun to have interesting problems as well. >> Yep. Yep. Thank you. >> So, we've talked a lot about how using agents is central to your workflow, but but this is an emerging technology. It's not established. There are many problems. So, what sort of complications does this bring? >> Well, uh I actually believe that uh Versal has to be good for everything, like regardless if it's in the AI agent or just like a complicated complicated piece of uh enterprise software. And what ultimately like joins all of those things brings all those things things together is that agents are dynamic. They need to stream their results. You need to basically be able to very quickly uh like react to LLM messages and re-render them in the browser. Everything has to be fluent and everything has to has to feel great. Uh and just building infrastructure for that to happen and for that to happen at scale and uh so that it like runs across across the entire planet essentially. That is actually quite challenging and this is again what Vercel Vercel is ultimately after. >> I really like how LLMs have turned from a machine learning problem to an engineering problem. I think like Claude Code was really the first one to break the back of it. But like obviously all of this stuff has been an application since ChatGPT became more than just a chatbot. So yeah, it's again, it's exciting to see the direction it's moving in. >> Yeah, absolutely absolutely. I'm excited as well. >> So [snorts] we talked about how the Gel team was acquired last year. So what specifically was behind the acquisition and what are you working on right now? >> So uh we joined Vercel we joined Vercel and our first task essentially was to analyze where the gaps are like uh what are the gaps with uh the existing Python support back in the time. Uh and uh obviously we're fixing the first problem which is you have a standalone Python application and how do you even deploy that to Vercel and make sure that all the dependencies are installed and the run time is primed to run your application and stuff like that. Fixing here and there some DX uh edge cases. >> Mhm. >> Uh but the pretty much immediately we realized that the biggest unlock is going to be understanding how people typically package front end and back end together because like sure, there is a bunch of Python applications that are just Python, but there is also a lot of Python applications that are Python and JavaScript because they render something in the browser. How do we make DX for that better? And uh Vercel DX is uh is is actually quite amazing for JavaScript. Like for example, you have a Next.js application, uh you create a new branch, you push that branch to GitHub, and you have a preview deployment, you can click on it, you can test it, you can comment in the preview. There is a lot of automation, a lot of UI, a lot of uh developer tooling built around enhancing your development workflow. But it wasn't quite working for um applications that are hybrid, that have a Python component and JavaScript component. So, the first task for us was to actually fix that. And just recently, we announced Vercel Services, which is a way uh for you to package your Python back end and uh let's say Next.js or like any JavaScript front end, potentially more components. You can You might also have like microservices in Go and Rust. How can package all of that to one single Vercel project and get the same experience. You run VC Vercel Dev command locally, and the whole development environment is spin up for you. So, you don't need to have like multiple terminal windows or like mess with Docker Compose. We just like read your code, automatically configure everything for you, recognize the frameworks that you code, and um everything is wired up for you. So, run Vercel Dev locally, it works. You run VC deploy, and it deploys to the cloud. Uh and everything is automated, everything is connected. So, like just building that DX was like the biggest challenge for my team, and uh we have a lot of interesting stuff ahead of us. Uh the compute team of Vercel with some collaboration with us shipped uh Vercel containers. So, now you can deploy Docker containers to Vercel. So, if you are not interested in serverless, you can deploy your Docker to Vercel. And it's also going to be all connected with the front end you will have all the beautiful AI workflows and tooling at your disposal. And then the other thing that we're building right now is Python AI SDK to have some counterweight for the JavaScript AI Vercel SDK. And that is also quite an exciting project. >> How big is your team? It's a lot of projects going on. >> My team is 12 people or maybe 11 people. I'm not too much focusing on the number. But we are we are quite productive, and I think that's that's that in general applies to Vercel just like an incredibly dynamic and active company within like we we we do a lot of things per capita. >> I I can say that and on the cutting edge as well. >> Yes, exactly. >> documentation. >> Yeah, exactly. Of course. Of course. Of course. >> Yes. >> So obviously you have a extremely deep history in Python. We've already talked about that a little bit. So would you mind going into some of the projects that you've worked on in Python and how they impact the work you do at Vercel? >> Is this about like previous projects or >> Yeah, so sort of the like the packaging work you've done in Python. >> Okay. Okay. >> Open source contributions, things like that. >> Right. Right. So again like my most extensive contributions are probably around async await world. Part of Python. UV loop is like one of the like major things. I think FastAPI Well, UV loop itself has been downloaded more than billion times now. This is like an implicit dependency of FastAPI. FastAPI depends on UV loop. So like if you run FastAPI in production, you basically run it on top of on top of that thing. And I have some plans for UV loop. Like I have plans to potentially write it in Rust to make make even little much to do some interesting tricks with HTTP, and uh I know that Vercel is quite open to to to sponsor this work and like let us do it. >> Mhm. >> Before we do it, first order of business is make sure that the platform is ready, that the that that that that Python is supported, and everything else. But like this is my dream to actually make it look even more capable as soon as we have a little bit more resources. Uh but now my async await past is strangely connected to the AI SDK future for Python because uh Python AI SDK that we're building is incredibly heavily async await built. Like everything is about messaging, everything is about streams, everything is about async await and asynchronous generators. It's actually quite quite interesting in that regard. I encourage you guys to check it out. Um so this is where my notion of async IO was like directly applied, like figuring out how can we like make it like really ergonomic and like really smooth. >> I was thinking this when I was researching for the interview because the amount of data coming through, obviously you're going to like you're doing a lot of IO operations, and so obviously this is going to be incredibly relevant. Yeah. >> Yeah, amount of data is 100% important, but also it's the reasoning, like understanding of you like how is the specific tool called by the agent? How can you make How can you introduce a sub agent that is also streaming? And what if that sub agent has tools that are also streaming? And now you need to join those streams and forward them to the browser so that the UI can render all of that and render it consistently, and ideally with like as low latency as possible. So it's it's it's it's stuff like that, and I think this is ultimately where async await is best at. Like I I can't even imagine how you would do it without async await, which is like threads and standard synchronous programming. So again, like I think that async await was kind of like it had to exist >> Yes. >> for us to build agents in the future. >> It's true, actually. You're like, good thing I did that. >> Yeah, exactly. [laughter] >> So, obviously we've talked about the fact that you contributed a lot to Python and you're talking about, you know, you really want to make time to do that and Vercel is supporting it, which is amazing to hear. So, in terms of the kind of cutting-edge problems that you're dealing with right now, where do you see gaps in Python and where do you think you might be able to contribute to that? >> So, um there's probably a lot of gaps, but I can talk about a specific gap that I'm actually working on with um Michael Sullivan, who is my PyTorch developer. Uh we are introducing a new PEP, Python Enhancement Proposal PEP 827, uh which is PEP 827, type manipulation. Essentially, that proposal allows us to have programmatic types, so that you can derive your types from other types programmatically, and type checkers will be able to understand that. That actually unlocks a lot of interesting uh things for Python, because it allows you to match Python dynamism like metaclasses and uh dynamic attributes and all of that to the type system. Because right now, type system of Python is relatively rigid and relatively restricted, so uh you cannot type Python accordingly. And uh what this boils down to is that if you want to use strict typing and let's say create a simple fast API product application, then suddenly you need to copy-paste data classes and stuff like that. Like imagine that you have an endpoint that uh requires um that creates a user. So, that endpoint has uh uh an input that is structured as data class, let's say username required and uh email required, both as strings. So, so far, so good. But then you want to have an update uh uh endpoint. And for that uh update endpoint, both of them must be optional. So, currently what you have to do, you have to just copy-paste. Now you have two data classes that are essentially the same, but the second one for both of the fields, they are basically or not, like they're optional. Uh and I think that is completely pointless exercise because it's just like balloons the amount of code that you have to uh that you have to ship, run, uh support. Uh this can be done automatically and this is done automatically in TypeScript, like for example in TypeScript you can express uh a type that takes a type and makes all of its fields optional or some of its fields optional. This is like a very naive example, but it sort of like shows the power of what you could be uh doing if you had that capability. And that PEP 827 specifically focused on that. And I think it's it's going to be a big unlock uh for uh for Python engineers in general, but also for Python engineers uh working engineers who actually use a generic coding because like agents are pretty slightly less cool for you to review, slightly less cool for you to reason about. >> Yep. >> Uh which is I think incredibly important. >> It's also it's so important because of the flow through agents is so um it's so important to have the type clearly defined. And so yeah, it's it's super interesting to hear about. >> I'm super excited. Like we initially started working on this uh back at Gel because what we wanted to have is an ability for you to express your database query in Python code and the result of the database query to be like inferred automatically just like from the shape of your code and how it uses the API, which is completely impossible right now in Python. >> Mhm. >> Uh so we started working on that and then when Vercel acquired us uh and we explained, "Okay, like we uh we we started building this. This has a lot of potential." Uh Guillermo of Vercel says, "Yes, let's absolutely finish the work and let's ship it." >> Yeah. >> And uh we we spent quite a lot of time uh making this proposal happen. Uh take a look at it. Like we we forked my pie and there is a my pie that supports all of this incredibly advanced type manipulation. We can do this in run time automatically so that like your identity models can compute the types and and it plugs there neatly as well. It's quite a feat. Obviously it's going to be a huge uphill battle for us to push this through and get this proposal approved, but we're working on that. So, let's see. >> It's it's also very nice to hear about this work when we have all this conversation about, you know, how AI is basically replacing developers and like this story is so different and like this is the real story. The story is that Python is growing. The Python is the language of AI is growing with AI. We've got people like you who are pushing the limits and taking us to the next steps. >> Thank you. >> So, Vercel recently became a PSF maintainer and you're also sponsoring a core developer. Thank you so much. And now you're a platinum sponsor or one of the platinum sponsors for Euro Python and we really really thank you in the community for this work, but it would be really nice to hear what's the thinking behind it. Like why are you investing in Python at a time when a lot of people are not? >> Well, first of all, I think a lot of people owe to Python big time. They should invest. I will not be naming some of the biggest companies of the world that are not here and they should totally be here. And I think it's a job of pretty much every one of us who are working for bigger companies to bring that simple fact to their attention is that like they owe a lot of their success to Python and it's a very good idea to dedicate a tiny chunk potentially pocket change of their revenue to support this language and this community. And I'm incredibly excited that like we we're doing that at Vercel. For me, first of all, it's it's personal. Uh, I fundamentally and deeply believe that Python needs to be supported way more. Like for the effect that it has on this world, the amount of sponsors or sponsorship should actually increase. And uh, I'm I'm happy to be part of it. Second, uh, I don't want anyone to think that hey, Vercel will just like start supporting Python and uh, shipping some things, but they don't care about the community. This cannot be uh, this this is just like that too. We actually do care about open source open source is part of DNA uh, our DNA. I think like most of Vercel frameworks and uh, and technologies are fully open source under permissive licenses. Uh, we support open source deeply and we want to do exactly the same for Python. And uh, that's that's just the way to go. Like you do the open source work and you support community and you do financial support. All of those things are important and uh, important to be done. We are sponsoring Serhiy Storchaka Serhiy Storchaka who is developer in residence. I'm super proud of that fact. He is an absolute machine. Uh, I joined well, I became Python core developer in 2013 and I think he was already active at that time and he's been incredibly active since then. Uh, it's absolutely just insane amount of work that he is pushing uh, and he's been doing that for years even prior even before LLMs. >> So, while you're here, what is Vercel hoping to contribute to the Euro Python community? >> Honestly, like we're just happy that we're here. Uh, and I'm happy that we're here. Uh, I I just want to talk to people. I just want to talk to users. I want our team to be more aware of what people have what kind of problems people have for Python. Like that is the exciting part for me. Like to understand what struggles people have and like ideally to solve them. So, that's that's kind of the whole point. And uh, the second like smaller part of it is just like make people more aware that hey, Vercel is now investing into Python. We're trying to be the best at Python. We just want to make sure that we are we're heard this way. But ultimately to me like just like talking to people is more important to me than that. >> Yeah. Just being here and showing your face, supporting the community. And I think this interview is important, too. Telling the story. It's a >> Thank you for that, by the way. >> Yeah. I think so. Yeah. So, this is something we've touched on already, but Vercel has talked about building on open foundations. So, what does the Python community, not just the EuroPython community, but the whole Python community mean to Vercel, especially because you have this sort of JavaScript, TypeScript foundation. Where does Python fit in and where does the Python community fit in? >> Well, first of all, like we we are continuing to ship new open source things now for Python, too. Like Python AI SDK wouldn't happen without Vercel. Like we started building it at Vercel and we will continue building it at Vercel. And it's like Apache 2 license open source project. Actually, it's an I'm about to actually give a talk about Python AI SDK and on one of the slides I say "Humanity's last organic framework." >> [laughter] >> We spend a lot of time making sure that it's actually nice, that like the APIs are nice, that it's not We didn't just like throw the agent creator a new AI framework. No, like we genuinely want to improve it. We genuinely want it to be state of the art and it required a lot of iteration, a lot of thought, a lot of fighting in front of a whiteboard to make sure that it's like a cohesive, composable, and like beautiful to to to use. So, we have this passion again in our DNA to not ship slop, >> [laughter] >> but like ship good stuff. And that will continue That will continue itself. Um Second what it means for Python, I think like I'm personally excited in in in in in part of the story that probably not too many people are even aware of and they also have some kind kind of Stockholm syndrome and they don't even know that it's a problem, but it is. Uh which is actually again going back in the conversation, which is the problem of how do you host your fast API application and your UI? >> Mhm. >> How do you how do you make that happen? Like it's weird to me that we that we are even having this conversation in 2026, but like it's still a problem and it's a multi-faceted problem because like how do you how do you do this locally? How do you deploy how do you deploy it? How do you observe it? How do you make sure that you can optimize it and make it look make it work faster? Uh and this problem is still unsolved. How do you make your Python code like be scalable? How do you integrate that into the into the uh dev tools and everything else? How do you give your agent your cursor agent access to all of that so that it can fix things for you uh uh knowing and seeing the big picture? It's it's questions like this and I wish they were answered, but they they they they are not and uh we're building that and by building that we will also be advancing the whole ecosystem forward. Like even the Vercel dev command that I was talking about that will launch the uh let's say fast API or Flask back end for you and your front end for you and join the the terminal output and like manage all those background services and commands just so that like it's cohesive. One command you run it and you can debug your application. I think that is already like step forward to the right direction and like I'm I'm happy if other frameworks start copying it. Like if fast API ships something just like that so that it's native to to their experience and some other Python framework does that and maybe some other JavaScript framework does that. So, it's like it's a unique opportunity for us to be thought leaders in how things like this should be done. I think like everybody will benefit. It's like such a huge win-win thing. Like Vercel as a company benefits if it ships genuinely good solution, people genuinely enjoy using it. The ecosystem wins because somebody shipped something good, so it can be copied and improved even further. So, win-win. >> Yeah, and focusing on proper architectural practices like proper software engineering practices. >> Exactly. >> Mhm. >> It's very important. >> It is very important. So, the final thing I wanted to ask you about is we've talked a lot about how Vercel is supporting the Python community and we love it. Um so, I guess just one kind of practical question. So, in terms of the work that you do for open source for Python, how much of it are you doing on on Vercel time versus your own time? >> Well, I wouldn't say that like I'm doing a lot of this work. Like PEP 8 to 7, the type manipulation PEP that we are working on now, uh like yes, that that is relatively involved. Maybe less uh for me personally, but like for example, Sally who is in my team, uh he spent quite some time to figure to figure it out and we'll continue to do so. Uh I personally I'm thinking about actually ramping up my contributions to Python and I'm pretty sure that Vercel is going to be supported because like a lot of uh the things that like I'm thinking about fixing or adjusting uh in Python are ultimately related to my work experience at Vercel. I see some problems and uh we'll eventually need to fix them. And like what better way of fixing them to just like go and fix Vercel. So, that's the goal. >> So, it's it's just it's really nice to hear about that and and really nice to yeah, just see a company taking from Python and giving back and we grow together, right? >> Yeah, thank you so much. >> Yeah. So, thank you so much. Like this was such a wonderful interview. It was such a breath of fresh air and um >> You're flattering me. >> [laughter] >> No, I think I've just really enjoyed it. So, >> Likewise. Likewise. >> Thank you for yourself for being with us at at EuroPython. Thank you for contributing back to the community in so many different ways. And yeah, thank you so much for giving us your time to tell us about your amazing career in Python. >> Yeah, exciting times. Thank you so much. >> [music]