About
I am a Software Engineer in Stripe's Machine Learning Foundations organization. I focus…
Activity
4K followers
-
Yuan Liu reposted thisYuan Liu reposted this📣 We are continuing expanding our teams in Toronto, Canada! Looking for a seasoned engineering manager and mid/senior level software engineers to work on Stripe's ML infra including large-scale feature platform, observability platform, LLM alignment, and more. We have ambitious goals to take AI/ML to the next level at Stripe. Feel free to ping me or hiring managers Uday Kumar Bandaru and Lei Zhang to learn more!
-
Yuan Liu reposted thisYuan Liu reposted thisWe just posted Stripe's 2025 annual letter: https://lnkd.in/gz2A6ztC.
-
Yuan Liu reposted thisYuan Liu reposted thisCalling all PhDs: Stripe's ML intern and new grad roles for 2026 just opened! Our interns and new grads have outsized impact across both research and product. A handful of highlights from this summer: 💡 Built an LLM-fine-tuned system for narrative generation with hallucination checks and LLM-as-a-judge scoring—now powering Stripe Smart Disputes. 🔍 Extended the sequence-to-sequence Stripe Foundation Model for generalized anomaly detection—and in the process discovered scaling laws for payment sequence models that reshaped our training approach. ⚡ Boosted fraud and retry model performance (and efficiency!) by training big, serving small: ensembling large models for maximum accuracy, then distilling them into a single efficient production model. 📦 Implemented a vector quantizer for the Stripe Foundation Model that cut storage and bandwidth costs by 200× at constant quality—then used those quantized embeddings to ship a detector for malformed payment data. 📉 Reduced token usage by 43% and operating costs by 80% in the fraud-review LLM pipeline while increasing recall by 5pp. Then extended to LLM distillation with a domain-specialized tokenizer and fine-tuned small model, shrinking context length by ~45% and beating baselines on precision. Ready to push forward research, ship to production, and shape products used by millions of businesses worldwide? Apply to join us via the links below. 🚀
-
Yuan Liu reposted thisYuan Liu reposted thisStripe isn’t a standalone company, we partner with some of the most influential companies to build what we build. Come help us with these partnerships around AI and future looking products. https://lnkd.in/gWxWb34Z
-
Yuan Liu reposted thisYuan Liu reposted thisWe’ve been using AI to help businesses on Stripe grow for more than 10 years. Here’s the latest: card testing attacks are down 80%, we recovered $6 billion in falsely declined transactions last year, and we're seeing 20% fewer unnecessary auth challenges.
-
Yuan Liu shared thisSuper excited to introduce Shepherd, Stripe's new ML Feature Platform based on Airbnb's Chronon. We truly appreciate the opportunity to collaborate with Suqiang Song, Nikhil Simha, Varant Zanoyan, Haichun Chen, Henry Saputra and the entire Airbnb team. We are thrilled to help open source Chronon and contribute back to the community.Yuan Liu shared thisI am super excited to share my team's work on the Shepherd project! A heartfelt thank you goes out to Yuan Liu, Kumar Chellapilla, and Vladimir Zhukov for their unmatched leadership support and invaluable guidance. I also want to give a shoutout to Daniel Kristjansson, Benjamin Mears, Piyush Narang, and Jeremy Robin for their excellent technical leadership. Huge Kudos to the AirBnB Team!!! Our collaboration was truly amazing. Check out this blog post to learn more about how we adapted Chronon to scale ML feature development: https://lnkd.in/gnzx6prv. #teamwork #collaboration #ml-infrastructure #feature-engineering #chrononShepherd: How Stripe adapted Chronon to scale ML feature developmentShepherd: How Stripe adapted Chronon to scale ML feature development
-
Yuan Liu reposted thisYuan Liu reposted thisAfter spending the last year working on Stripe's AI/ML product suite, I'm excited to share that we're growing our Product team! We have a huge opportunity to help businesses to grow even faster with AI/ML powered features, and are looking for a PM familiar with ML that can help us shape experiences for the millions of businesses that use Stripe:
-
Yuan Liu reposted thisYuan Liu reposted this🚀 Exciting News from the Tech Front: Chronon is Now Open Source! 🚀 After 20 months of intense collaboration, sweat, and occasional tears, our MLI teams at Airbnb, Uber, and Stripe have finally made the Chronon project open source. This journey began back in 2017 with its predecessor, Zipline. 🔍 Chronon vs. The Hype: Unlike projects that explode overnight on the generative AI hype wave, Chronon promises to be a steady flame, illuminating the ML feature development and deployment path. It's the fruit of joint efforts by Silicon Valley's leading tech firms, driven by a passion for innovation and open-source collaboration. ✨ The Journey from Zipline to Chronon: What started as Zipline had to navigate the tumultuous waters of trademark disputes, emerging stronger with a refined focus on real-time, streaming, and batch processing for ML features. This transformation was about overcoming technical hurdles and fostering a culture of openness and collaboration within and beyond Airbnb. 🌉 Building Bridges: Our first milestone was internal buy-in, turning technical challenges into opportunities for cross-departmental collaboration. The subsequent milestone involved forging partnerships with Uber and Stripe, showcasing the Silicon Valley spirit of open innovation. Our collaboration with Uber aimed to integrate Michelangelo and Chronon, setting the stage for a shared MLOps ecosystem. 💪 Stripe's Stellar Contribution: Stripe's involvement with Chronon has been transformative. Tackling massive throughputs and data processing challenges, Stripe's team played a crucial role in enhancing Chronon's technical quality and maturity. Their strategic investment in Chronon's development underscores the project's potential to shape the future of ML feature platforms. 🎉 Celebrating Collaboration: Our journey with Chronon has been a testament to what can be achieved through collaborative innovation. From technical forums to strategic partnerships, we've laid a solid foundation for Chronon's growth and its role in advancing AI technologies. 🌟 Support Chronon: Join us in this exciting phase by starring the Chronon project on GitHub and diving deeper into our journey through our engineering blog. 🔗 GitHub: https://lnkd.in/gsTq6kUW 🔗 Engineering Blog: https://lnkd.in/gAqPaCtQ #OpenSource #Innovation #MachineLearning #Chronon #Collaboration #TechNews
-
Yuan Liu reposted thisBringing the magic of Chronon to Stripe is the project I've been DRI for since shortly after I joined the company. Having access to the code let us adapt it to meet our feature freshness needs and let us swap in K-V stores which made sense in our context. Now you can too!Yuan Liu reposted thisWe just open sourced Chronon - Airbnb’s ML Feature Platform! (If you're interested in chatting with us about it, feel free to join our Discord: https://lnkd.in/gAVb95XW). We started working on this project seven years ago, when we observed a simple but big problem — ML practitioners weren’t spending that much of their time on ML. Rather, they were bogged down with the huge amount of “plumbing and gluing” that it takes to actually get their models into production. As we dug deeper into this problem, we identified feature/data engineering as a major bottleneck in the whole process. Streaming feature computation, scalable backfills, windowed aggregations, and low latency online serving are all complex infrastructure challenges that need to come together to make an effective workflow for model development and productionization — not to mention observability and governance. It’s unreasonable and inefficient to ask the ML practitioners to think about all of this, so we set out to build a platform that abstracts away this complexity. ML practitioners simply define their features in our DSL, and all of the data plumbing, backfilling, and serving is handled for them. After rolling out the platform within Airbnb (it now powers all major ML models within the company), we started thinking about open source. We were especially lucky to have met the Feature Platform team at Stripe, who became our partners in the open source effort! Huge thank you to the teams at Airbnb and Stripe (https://lnkd.in/garNHyZD) for excellent work behind this! https://lnkd.in/guUZjaQHChronon, Airbnb’s ML feature platform, is now open sourceChronon, Airbnb’s ML feature platform, is now open source
Yuan Liu commented on a post
5mo
Love the thesis and also agree. To push it a bit further, what about the revenge of the PM? Test-driven development feels like both finally feasible and also a must-do with AI coding, but who is going to review the tests? Maybe we need spec-driven development and have human review the spec. AI can then review the test against the spec, while the (deterministic) computer can verify the code against the test.
-
Yuan Liu liked thisYuan Liu liked thisThis week is my last at Stripe. I joined in 2018 to work on Developer Productivity, focused on helping Stripe’s engineers have the most productive time of their careers. That started with the inner loop, trying to make the developer experience as tightly integrated as possible. I truly believe that Stripe has some of the best developer tools in the industry. I’m incredibly proud of what we built, but much more grateful for the people I got to build it with. Stripe is genuinely full of people who care deeply about their work, hold themselves and each other to a high bar and are just great people to spend your days working alongside. I’ve learnt an enormous amount from them, especially my managers Aaron Spinks, Rahul Patil and Will Larson. Thank you to all the current and former Stripes who made the last eight years so memorable. I feel incredibly fortunate to have spent them here.
-
Yuan Liu liked thisYuan Liu liked thisExcited to share I've joined Google Cloud, where I'll be building agents that can run your cloud end-to-end! I loved helping developers get the most out of frontier models via the OpenAI API, but couldn't pass up the opportunity to work at the Agent layer again with Ryan Lopopolo. If you're excited to learn more and get to work with folks like Hemal Shah, Henry Scott-Green, Alexandre R. and more - please reach out!
-
Yuan Liu liked thisYuan Liu liked thisAfter 8.5 years of working on Developer Productivity at Stripe (and hopefully making the developers more productive) I have decided that the next step for my personal development is to become deeply unproductive for a while. No plans in the short term but extremely excited to take a break before deciding what's next. A massive thank you to Scott MacVicar, Yvette Nameth and everyone else I had a chance to work with who made the time at Stripe as fun as it was!
-
Yuan Liu liked thisYuan Liu liked thisEveryone's talking about loops and factories and token maxxing. This is all easy if you have a demo app or a side project. It becomes challenging when you have a production system with dozens of engineers. A scaling business typically has complex infrastructure -- many services, databases, security keys, AWS dependencies, compliance requirements. The list goes on. There's no closing the loop if your agent can't access all that infrastructure. You're wasting tokens and more importantly slowing down product velocity. The first piece of the puzzle is to build out your agent dev environment. One way to think of this is that your agent should be able to spin up a scaled replica of the production system with all its nuance and complexity. This can take months to build and a team of engineers to maintain. The second piece is to give all frontier agents access to this infra. You don't want to just support frontier models. You want support frontier coding agents in their raw glory. This is important because frontier models are trained almost entirely on their own harness traces and will work best with their own harness in the long run. If you're on subscription plans, then supporting the harness keeps your costs down too. The final piece is being able to allow humans to seamlessly take over when the agent gets stuck. If you have to download a big PR to a laptop and reproduce the agent session, then all those agent tokens have been wasted. The right UX should feel like the agent walked up to engineer's desk with their laptop to ask for help. I've brainstormed how to set this up with dozens of companies. Let me know if you'd like to chat about your agent dev infra!
-
Yuan Liu liked thisYuan Liu liked thisCan you design caching algorithms that are both simpler and better than decades-old classics? The answer is yes—but only with the right data. The video recording for a talk I gave at South Bay Systems (https://lnkd.in/gtigjNYD) about how we found such algorithms after mining the largest collection of production cache traces is now available: https://lnkd.in/gCFCNxjK As promised, the slides for this talk are now available too: https://lnkd.in/gy5jqDwu I cleaned up the transcript generated by YouTube (which is very bad at jargon) and put it into the presenter notes on each page for reference. Most of the work in the slides are part of Juncheng Yang's PhD thesis, which won the 2025 ACM SIGOPS Dennis M. Ritchie Award. You should totally give him your production traces (e.g. from LLM inference) so he and his students can continue to produce amazing work that you can put into your systems in the end :-) Below is the last slide that I didn't have time to cover—it is a synthesized map of cache research lineage that led us to the designs in the talk. It's very satisfying to come up with designs so simple and yet powerful decades after the original ideas took shape.
-
Yuan Liu liked thisYuan Liu liked thisAt 11k employees, our AI costs are going up. Which model & harness should we use to lower cost but also retain great quality? We didn't want to blindly trust public benchmarks. So we ran a comprehensive evaluation on our tasks, code base, infra. It's been produced by more than 3,000 software engineers, spans 3 hyperscalar clouds and many languages and tasks. The results are surprising. We find that for the SAME mdoel, the choice of harness can significantly save costs (~2x). We also find that GLM 5.2 performs extremely well. We run 𝐎𝐦𝐧𝐢𝐠𝐞𝐧𝐭 in front of these and can easily multiplex different harnesses and models for different tasks. Check it out: https://lnkd.in/gtmRvgng
-
Yuan Liu liked thisYuan Liu liked thisLast year, I went viral for spending $9k in a month on Cursor. One year later, I'm spending over $2k a day on a mixture of agents ($3k on my best day). I've shipped 301 PRs in 12 days (44 yesterday, 78 on my best day). And the charts are going up. I gave a bunch of talks on "Velocity Coding", my approach for building with coding agents at breakneck speed without sacrificing quality. Might be time to refresh that content for 2026. Comment below if you're interested in a webinar!
-
Yuan Liu liked thisYuan Liu liked thisI published a new IronCurtain release. The theme is about making agent runs repeatable, auditable, and useful as training data. What shipped: - Codex CLI adapter with faithful token-trajectory capture, so agent runs can become training data for SFT and RL. - A new evolve workflow for iterative candidate evolution: deterministic container execution, structured verdict routing, human gates; based on ASI Evolve. - Snapshot and resume for shared containers across stops, plus runtime dependency install instead of per-workflow baked images. - Cargo/crates io support in the MITM package proxy, extending install mediation past npm, PyPI, and apt. - Memory MCP server 0.2.0 with atomic-fact ingest and parent-context retrieval. - Vulnerability-discovery hardening against masked findings and hanging harnesses. IronCurtain is open source on GitHub: [Link censored by LinkedIn]
-
Yuan Liu liked thisYuan Liu liked thisAfter four wonderful years at Stripe, I'm thrilled to share that I've joined the Finance & Strategy team at Anthropic. I owe Stripe so much and I will always carry gratitude for the amazing people I had the privilege of working with. They taught me the real meaning of executing well with grace and skill under pressure. Could not be more excited for what's ahead. Getting to do finance and strategy at a company building frontier AI, with this much care and seriousness about doing it well, is the kind of work I've hoped to do for a long time. I'm honored to be here and can't wait to dig in with this team. To everyone at Stripe, thank you and to everyone at Anthropic, so glad to be here!
Experience & Education
Publications
Languages
-
English
-
-
Chinese
-
Other similar profiles
Explore more posts
-
Maysam Sadeghi
The Claude ecosystem is exploding faster than I expected. My newest favorite plugins: - Get Shit Done (orchestration layer that plans and builds apps phase by phase) - Ralph Wiggum Plugin ( iterative, self-referential AI development loops) - n8n-MCP (direct Claude integration with 1,000+ automation nodes) What's interesting is how quickly we've moved from simple AI coding assistants to full orchestration platforms. The shift feels similar to what happened with CI/CD tools a few years ago. First you had basic scripts, then comprehensive pipelines, now full DevOps orchestration. We're seeing the same evolution with AI coding tools, just compressed into months instead of years. #AI #Claude #SoftwareDevelopment #AIAgents #DevTools #Automation
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content