Activity
2K followers
-
Shunyao Li reposted thisShunyao Li reposted thisNew on the Pinterest Engineering Blog: Tracking Down Mysterious ML Training Stalls 🔍 In recent efforts to upgrade PyTorch's version as part of Pinterest's ML training platform, the team saw a sharp drop in performance. Discover their journey debugging the root cause of this issue in a new article written by Chen Yang, Shaochen Y., Shunyao Li, Chantat Eksombatchai and Mark Molinaro. https://lnkd.in/etbfJ3A3
-
Shunyao Li reposted thisShunyao Li reposted thisAn ML stack upgrade (e.g., PyTorch, CUDA) always takes longer than expected, and we learn new things every time we undertake one. Chen Yang Chantat Eksombatchai Shunyao Li Mark Molinaro Shaochen Y. and Charles-A. Francisco captured some of our learnings this time, and there are some insightful tips on profiling and perf debugging. https://lnkd.in/g_yj5GKY
-
Shunyao Li shared thisExcited to share our team’s new blog on PinCompute - A Kubernetes Backed General Purpose Compute Platform for Pinterest 📌Shunyao Li shared thisSuper excited to share our next generation Kubernetes backed #pinterest compute platform — PinCompute. Very proud of the team and our partners! https://lnkd.in/gAivAV6PPinCompute: A Kubernetes Backed General Purpose Compute Platform for PinterestPinCompute: A Kubernetes Backed General Purpose Compute Platform for Pinterest
-
Shunyao Li shared thisI'm thrilled to share my first Pinterest Engineering Blog article! At Pinterest, our mission is to bring everyone the #inspiration to create the life they love. To deliver this value, we prioritize the quality and availability of our work. Check out this post about how we improved our Kubernetes control plane performance! #pinterest #kubernetes99% to 99.9% SLO: High Performance Kubernetes Control Plane at Pinterest99% to 99.9% SLO: High Performance Kubernetes Control Plane at Pinterest
-
Shunyao Li liked thisHonored to deliver the keynote speak in this year's PyTorch Conference NA. I will share the most recent updates from DeepSpeed.ai team on scaling large model training. Kudos to the awesome leads for the great work in the past year (Masahiro Tanaka Tunji Ruwase Minjia Zhang Guokai Ma)Shunyao Li liked this🎤 Meet Zhipeng Wang, PhD, Senior Staff Software Engineer at Google and keynote speaker at PyTorch Conference North America. https://lnkd.in/gfDpwPxA In his keynote, Scaling Large Frontier Models Training with DeepSpeed, Zhipeng will share the latest technical advances from the DeepSpeed team, including model-systems co-design and optimizations for diverse architectures, workloads, and accelerator platforms. He will also explore how collaboration among model developers, hardware vendors, and AI infrastructure partners is expanding the DeepSpeed open source community and supporting emerging large-scale AI workloads. 📅 October 20-21 📍 San Jose, California Register to join us: https://lnkd.in/gKzzmr_7 #PyTorchCon #PyTorch #PyTorchFoundation #FutureOfAI #AI #GenAI #MachineLearning #ML #DeepLearning #OpenSource #OpenSourceSoftware #OpenSourceDevelopment #OpenSourceCommunity #OSS #LinuxFoundation #events #linux
-
Shunyao Li liked thisShunyao Li liked this🚀 We’re Hiring: Sr. Manager – AI/ML and more! My team is looking for a Sr. Manager to lead our core innovation in recommendation systems, focusing on generative recommendation, LLM4rec, foundation models, and reinforcement learning. If you want to lead world-class ML talent, bridge cutting-edge research with massive production impact, and push the boundaries of AI/ML, we’d love to connect! 📄 Recent team publications: • UniPinRec (RecSys 2026): https://lnkd.in/gq4c5E9Y • PinRec (KDD 2026): https://lnkd.in/gqMsFQ4m • PinFM (RecSys 2025): https://lnkd.in/g8igYTdf 👉 Apply or DM me: • Sr. Manager Role: https://lnkd.in/gxbNhCjd • IC Roles: https://lnkd.in/gXdfFVYpUniPinRec: Unifying Generative Retrieval and Ranking at Pinterest ScaleUniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale
-
Shunyao Li liked thisShunyao Li liked thisBrother, can you spare a token? Prudence is important in our age of tokenmaxxing but that’s not the only reason we built Pinterest’s cloud-hosted VLM serving stack with NVIDIA Dynamo and Blackwell GPUs. We wanted a platform which could convert rich visual context to reusable embeddings to dramatically speed up real-time multi-modal AI while giving our teams a common foundation to ship new experiences like Pinterest Assistant, hybrid search and more. This was the result of close collaboration with NVIDIA and is also one of the very first use cases for PinCompute EKS, Pinterest’s next gen compute infrastructure. Many thanks to my co-authors, Lei Pan, Salina Wu, Cristian Lopez, Guantong Bai, Saurabh Vishwas Joshi, Chia-Wei Chen, Ambud S.en, Ambud S. and the incredible Pinterest and NVIDIA teams that made this possible! https://lnkd.in/gWNkXyeC cc: Elijah Soba, Qi Wang, Anthony Casagrande, Guan Luo, Kris(Yu-Hsin) Hung, Ryan McCormick, Harry Kim, Akshatha Kamath, Matt Rawson
-
Shunyao Li reacted on thisShunyao Li reacted on thisHad a great time speaking at Ray Summit 2026 about The Synthetic Data Flywheel at Pinterest. We shared how we’re using Ray Data and vLLM to build large-scale synthetic data pipelines for post-training — combining GPU inference, tools, and LLM judges into scalable post-training workflows. Ray Data has been a great fit for these kinds of heterogeneous workloads, especially when dealing with stages that have very different scaling and concurrency characteristics. This work was truly a team effort. Huge thanks to Yash Upadhyay , Raphael Poulain, Tianjian Huang, Zhenyu Tan, Shubham Gupta, Shunyao Li, Luyan Wu, Korhan Citlak, Dave Chen, AI Training team at Pinterest! A special thank-you to Karthik A.ial thank-you to Karthik A. and Bo Liu for their leadership and support. Really enjoyed sharing what we’ve learned with the Ray community. Thanks to everyone who joined the session, to Eric W. for co-presenting this with me, and to Anyscale for putting together a great Ray Summit! #RaySummit #Ray #LLM #PostTraining
-
Shunyao Li liked thisShunyao Li liked thisTauGrid is public now: https://lnkd.in/gB3aN7sg TauGrid is an open-source, cloud-native AI infrastructure project for teams running GPU workloads on Kubernetes. The goal is simple: make it easier for platform teams and researchers to run, schedule, and monitor AI workloads without having to assemble every piece of the stack by hand. TauGrid brings together: * tau CLI for submitting and managing workloads * workload queueing and admission with Kueue * Ray cluster orchestration with KubeRay * GPU health monitoring and node diagnostics * observability for clusters, GPUs, and workloads It supports the full AI workload lifecycle: data preparation, distributed training, fine-tuning, and inference. This project is still early, but I’m excited that it is now public. AI infrastructure is moving fast, and I believe the next wave of useful AI systems will depend on boring-but-critical platform layers: scheduling, isolation, GPU fleet health, observability, reproducibility, and a clean developer experience. If you work on AI infrastructure, Kubernetes, GPUs, distributed training, Ray, scheduling, or platform engineering, I’d love your feedback. #OpenSource #AIInfrastructure #Kubernetes #GPU #Ray #Kueue #PlatformEngineering #TauGridGitHub - Azure/taugrid: Cloud-native AI infrastructure for teams to run, schedule, and monitor AI workloads on GPU-enabled Kubernetes clusters, from data preparation to distributed training, fine-tuning, and inference.GitHub - Azure/taugrid: Cloud-native AI infrastructure for teams to run, schedule, and monitor AI workloads on GPU-enabled Kubernetes clusters, from data preparation to distributed training, fine-tuning, and inference.
-
Shunyao Li liked thisShunyao Li liked thisPinterest Engineering made a strong mark at ACM SIGKDD & Annual KDD Conference 2026 in Jeju, Korea, showcasing the breadth and depth of our work. From papers and posters to workshops and conference chairs, Pinterest was a proud sponsor of this year's conference and saw tremendous traffic, curiosity, and excitement at our booth. Highlights of KDD included VP of Engineering, Faisal Farooq’s talk, “PINSAFE: A Framework for Building the Positive Corner of the Internet,” and Pinterest receiving the Best Paper Runner-Up Award in the Applied category for “Deep-learning Causal Retrieval Optimization for Efficient e-commerce Distribution in Pinterest.” A great week for our team and an exciting demonstration of how we’re advancing AI, large-scale systems, and the future of discovery by developing cutting edge AI technology. Explore our #KDD2026 papers below and apply to job openings at pinterestcareers.com. 📌 Deep-learning Causal Retrieval Optimization for Efficient e-commerce Distribution in Pinterest https://lnkd.in/eX9k2H8v 📌 PinEqualizer: Full Funnel Content Exploration and Debiasing System at Pinterest https://lnkd.in/ebmxjUjD 📌 PinRec: Unified Generative Retrieval for Pinterest Recommender Systems https://lnkd.in/eCEqtryj 📌 Pinterest Canvas: Large-Scale Image Generation at Pinterest https://lnkd.in/ebTkazB8
-
Shunyao Li liked thisShunyao Li liked thisYesterday was my last day as a SWE Intern at Pinterest! 📌 I am incredibly grateful to have had the opportunity work alongside such a talented and supportive team. During my time there, I learned a lot about backend systems, Kubernetes, and what it takes keep an AI training platform reliable all while ensuring a good user experience. Throughout the course of this internship, I have noticed my skills and confidence as a Software Engineer have increased. A massive thank you to my mentor Anna Brower for your amazing guidance and helping me ramp up so quickly, Mingyuan Tian for sharing your knowledge about Kubernetes, my manager Karthik A. and the entire ML training team for being so welcoming! Thank you also to my recruiter Destiney Williams and the entire University Recruiting team for this opportunity! I am exicted to take these learnings with me to the future! 🚀
-
Shunyao Li liked thisShunyao Li liked thisAfter 6+ years at Pinterest, yesterday was my last day. I joined the company seeking a new challenge mostly based in 1 time zone with an in-person culture, and I was thrilled to work at a company whose mission centered around bringing joy and inspiration to the world. I interviewed in Feb 2020 when there was “some health thing” going on that would surely be resolved before I joined, then was the company’s second fully virtual onboarding cohort in April 2020. I certainly experienced the challenge I was seeking, but what I didn’t know was how the mission would permeate to the people who worked here. I found a culture where people are genuinely invested in each other’s success, and willing to stretch themselves to offer their support and expertise to get to better outcomes and learn something new along the way. Thank you to everyone who took a chance on me, taught me something new, challenged me to think differently, helped me through hard moments, celebrated the wins, or made the day-to-day feel lighter and more fun. I’ve been lucky to work with people who care so much. I’ll be invested in the success of the team long after I can call myself a Pin-ployee, thank you Pinterest for being such a special chapter of my life! I’ll share more about what’s next soon.
Experience & Education
Courses
-
Algorithm Design and Analysis
-
-
Big Data Management Technology
-
-
Computer Architectures
-
-
Computer Networks (Honor Track)
-
-
Computer Organization
-
-
Data Structure and Algorithm
-
-
Introduction to Deep Learning 11-785
-
-
Operating Systems
-
-
Practice of Programming in C&C++
-
Projects
-
Pun-GAN: Generative Adversarial Network for Pun Generation (EMNLP 2019)
-
See project1. Tackled a major challenge in pun generation, the lack of large-scale pun corpus to guide supervised learning, by proposing a generative adversarial network (GAN) for pun generation using TensorFlow.
2. Introduced word sense disambiguate discriminator into pun generation task for the first time to generate vivid puns similar to real sentences.
3. Built a generator to produce pun sentences and a word sense disambiguate discriminator to distinguish between generated pun sentences and real…1. Tackled a major challenge in pun generation, the lack of large-scale pun corpus to guide supervised learning, by proposing a generative adversarial network (GAN) for pun generation using TensorFlow.
2. Introduced word sense disambiguate discriminator into pun generation task for the first time to generate vivid puns similar to real sentences.
3. Built a generator to produce pun sentences and a word sense disambiguate discriminator to distinguish between generated pun sentences and real sentences with specific word senses.
4. Released open-source code at https://github.com/lishunyao97/Pun-GAN.
Honors & Awards
-
Award for Excellent Scientific Research
Peking University
-
Outstanding Undergraduate Scholarship
Peking University
Languages
-
English
Professional working proficiency
-
Chinese
Native or bilingual proficiency
Other similar profiles
Explore more posts
-
Tan Wang
Over the past year, my team and I built Pinterest's Model Context Protocol (MCP) ecosystem from the ground up: a central registry, a growing fleet of domain-specific MCP servers, and production integrations across our IDEs, internal chat surfaces, and AI agents. MCP is an open standard that gives large language models a unified way to talk to tools and data sources. At Pinterest, we used it as the substrate for AI agents that can safely automate real engineering tasks, from querying Presto data on demand to diagnosing Spark job failures and surfacing institutional knowledge. A few things I'm particularly proud of: * We designed for security from day one: two-layer auth (end-user JWTs + mesh identities), business-group-based access gating, and mandatory human-in-the-loop for sensitive actions. * We made it easy for any team to ship a new MCP server by creating a unified deployment pipeline that handles all the infrastructure, so domain experts just write business logic. * The ecosystem has scaled to 66,000+ invocations per month across 844 monthly active users, saving an estimated 7,000 engineering hours per month. This was a cross-functional effort. Thank you to everyone on the Agent Foundations, Security Engineering, and Traffic Engineering teams who made this possible, and to our engineering sponsors for their support and guidance. Full write-up on the Pinterest Engineering Blog: https://lnkd.in/e-mH26Bw
7 Comments -
Vivian Di Bai
What happens to all the attempts an AI agent makes on the way to discovering a better solution? Dream-RSI keeps them as a world it can replay, asking what a different search would have found — no new experiments, no retraining. Glad to have been part of this work. Worth a read if you think about discovery loops and self-improvement in agents.
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content