Activity
2K followers
-
Mengjin Yan shared thisI’m hiring Staff Software Engineer to join Ray Core team at Anyscale! Ray Core is the foundation of the Ray ecosystem, powering libraries like Ray Train, Ray Data, and Ray Serve. With Ray used by companies like Nvidia, Uber, Spotify, and many others, the systems we build support AI workloads across the industry. If challenges like these excite you: → Designing the primitives behind distributed execution and scheduling → Improving resource utilization and performance across large clusters → Building fault tolerance for demanding AI workloads → Solving difficult concurrency, networking, and runtime performance problems …I’d love to connect. As a Staff Engineer, you’ll own a major technical area of Ray Core: identify the problems that matter, shape the roadmap and architecture, and lead projects across engineers and teams—all while staying hands-on with implementation and production debugging. What makes this opportunity exciting to me is the combination of technical depth, ownership, and open-source impact. Your work can help shape how developers across the AI ecosystem build and scale their applications. If your background is in databases, storage, streaming systems, schedulers, or distributed runtimes, this could be a great next challenge. Learn more and apply (https://lnkd.in/guYcU_CZ), or message me directly. Referrals are very welcome!
-
Mengjin Yan reposted thisMengjin Yan reposted thisEarly bird pricing for Ray Summit 2026 is extended through July 17. $200 for a Summit pass, $250 for Summit + Training, and rates go up after that. The full agenda is live, with teams from Google, Apple, Microsoft, Uber, Spotify, JPMorgan Chase, BMW, and many more. Banking, automotive, ride-hailing, drug discovery, mental health care, and more. Different industries, same scaling problems. San Francisco, Marriott Marquis, August 24–26. Register → https://lnkd.in/gqHdq__N
-
Mengjin Yan reposted thisMengjin Yan reposted thisThe Ray Core and Ray Data teams at Anyscale are actively hiring system engineers in Bengaluru! Ray Core serves as the cornerstone of the entire Ray ecosystem and powers libraries like Ray Train, Ray Data, and Ray Serve — quickly adopted by companies like OpenAI, DeepSeek, Spotify, Uber, DoorDash, Pinterest, Apple, and many more. Ray Data takes this further as the scalable data processing engine behind the training and inference pipelines of some of the largest AI models in the world. In this role, you will play a pivotal part in shaping the future of Ray and Anyscale. Particularly in the context of the growing importance of opensource and LLMs, you will be a crucial contributor to our strategic goal of establishing ourselves as the compute substrate of this unprecedented AI wave. If the prospect of tackling challenges like: → Scaling clusters to 10K+ nodes → Optimizing network transfer speed for petabyte-scale workloads → Building the data and execution infrastructure behind large multi-modal models → Pushing the limits of distributed training and inference …excites you, I encourage you to apply or message me directly if you have any questions. 📍 Bengaluru 🔗 Open roles: https://lnkd.in/ga6R6bvJ #Hiring #Bengaluru #DistributedSystems #RayCore #RayData #OpenSource #LLM #MLInfra
-
Mengjin Yan reposted thisThe addition of Ray marks another milestone in uniting the building blocks of open source AI under PyTorch Foundation. With PyTorch for model development, vLLM for inference, DeepSpeed for optimization, and now Ray for distributed execution, developers have a complete open ecosystem to build and scale AI responsibly. 🔗 Read Frederic Lardinois & The New Stack’s coverage: https://lnkd.in/gtx5Jdes? #PyTorch #OpenSourceAI #AIInfrastructure #DistributedAIMengjin Yan reposted thisThe Ray project will join existing projects like PyTorch itself, the vLLM inference engine and deep learning optimization library DeepSpeed. By Frederic Lardinois
-
Mengjin Yan shared thisLabel-based scheduling lets you place your Ray workloads on the right nodes with more flexibility. Congrats Ryan O'Leary, on the OSS contribution—and thanks to everyone across Anyscale and the GKE team at Google for making this happen! #Ray #Anyscale #GKE #raysummitMengjin Yan shared thisAnnouncing a powerful new feature in Ray that integrates it more seamlessly with Kubernetes and brings us one step closer to a serverless world. Excellent work by Mengjin Yan and Ryan O'Leary! This one was a collaboration between the Ray Core team at Anyscale and the GKE team at Google. I hope to see more big features coming out of this collaboration and others in the future. https://lnkd.in/gwE7HnzqIntroducing Label Selectors: Improved Scheduling Flexibility in RayIntroducing Label Selectors: Improved Scheduling Flexibility in Ray
-
Mengjin Yan reposted thisMengjin Yan reposted thisRay Summit is going to be excellent. Can't wait to hear from xAI, Perplexity, Cursor, Thinking Machines Lab, Physical Intelligence, Applied Intuition, Prime Intellect, vLLM, SGLang (sgl-project), and so many others. Some major themes this year that come up over and over - Reinforcement learning infra - Multimodal data (lots of video) - Distributed inference - Scalable agent infrastructure - Robotics / autonomy https://lnkd.in/gE8pn3sv
-
Mengjin Yan reposted thisMengjin Yan reposted thisThe Ray Summit 2025 agenda just dropped! 🔥 80+ deep technical sessions from Al researchers and builders across every industry from autonomous vehicles to finance and media. 👉 Check out the agenda → https://lnkd.in/e-_Sm2E3 This year’s focus: AI in production Hear from builders at Meta, Netflix, Apple, Cursor, Bridgewater Associates, J.P. Morgan, Adobe, Perplexity, Roblox, Anthropic, NVIDIA, Microsoft, Amazon, Google, Zoox, ByteDance, Coinbase, DataRobot, Autodesk, Grab, Pinterest, Character AI, Physical Intelligence, Applied Intuition and many more as they share how they’re building and scaling the next generation of distributed AI systems. 🗓 November 3–5 • San Francisco We hope to see you there!
-
Mengjin Yan reposted thisMengjin Yan reposted thisA huge fraction of the talks at Ray Summit will be deep dives into AI infra scaling, post-training infra, multimodal data pipelines, and agentic AI. The talks feature xAI, JP Morgan, Apple, Grab, Adobe, Applied Intuition, Character.ai, Physical Intelligence, Motional, and more. We saw some clear themes in the hundreds of submissions: a new variety of LLM inference challenges, new post-training scaling techniques, large-scale multimodal data pipelines, and agentic AI. Join me Nov 3-5 in San Francisco! https://lnkd.in/gE8pn3sv
-
Mengjin Yan reposted thisMengjin Yan reposted thisJust got a sneak peek of the breakout sessions lineup for #RaySummit2025 – and it’s 🔥 Sessions from: 🔹 Character.AI on Scaling LLM Post-Training 🔹 The State of vLLM in 2025 🔹 Roblox on Training 3D Foundation Models with Ray 🔹 xAI on Scaling Image + Video Processing 🔹 Zoox on Reliable, Multimodal LLM Serving 🔹 Perplexity on RDMA P2P for KvCache + Weight Transfer + MoE Looking forward to learning from the teams actually building these systems. Come join us. Save 25% with code ANYJOIN25 → https://lnkd.in/gjqHUc5i
-
Mengjin Yan liked thisMengjin Yan liked thisWe tracked 22 people who left LinkedIn, and the 17 companies they started. LinkedIn's engineers built Kafka, Pinot, and DataHub as internal tools first. The biggest became Confluent. IBM paid $11B for it in March. Here are 17 companies minted by the platform you're reading this on: 1/ Oscilar - Neha Narkhede Neha co-created Kafka at LinkedIn. Now an AI risk platform; MoneyGram picked it for global risk decisioning. 2/ StarTree - Kishore Gopalakrishna & Xiang Fu They built Apache Pinot at LinkedIn to power Who Viewed Your Profile. $75M raised to sell it to everyone. 3/ DataHub - Shirshanka Das & Swaroop Jagadish Shirshanka created DataHub, LinkedIn's metadata catalog. Now a company: $35M Series B led by Bessemer in 2025. 4/ Onehouse - Vinoth Chandar Ran LinkedIn's Voldemort database at millions of requests a second. His lakehouse startup has raised $68M. 5/ Esusu - Samir Goel Reports rent to credit bureaus. 272,000+ renters have gotten their first credit score. Valued at $1.2B. 6/ AgentSync - Jenn Knight & Niji Sabharwal They met at LinkedIn; Jenn was its first Salesforce engineer. Insurance licensing infra, $161M raised. 7/ Anomalo - Elliot Shmukler Helped grow LinkedIn from 20M to 200M members. Now: data quality monitoring, $82M+ raised. 8/ Merge - Gil Feig Gil built iOS at LinkedIn. Merge is one API for B2B integrations. OpenAI and Ramp run on it; $75M raised. 9/ Daffy - Adam Nash LinkedIn's VP of Product, then Wealthfront CEO. His donor-advised fund crossed $1B in assets this March. 10/ Cameo - Steven Galanis Sold LinkedIn's sales tools in Chicago before Cameo. 31,000 creators joined CameoX in its first 18 months. 11/ Rula - Gabe Diop From LinkedIn's sales program to mental health care: 23,000+ providers across all 50 states. 12/ Wisq - Jim Barnett & Chih-Po Wen Sold Glint to LinkedIn, stayed on inside, then left to build Harper, an AI HR generalist. $55M raised. 13/ Magical - Harpaul Sambhi LinkedIn bought his first startup, Careerify. His second automates healthcare admin. $41M raised, Series A led by Coatue. 14/ Neura Health - Elizabeth Burstein Senior PM at LinkedIn, now virtual neurology. 43,000+ patients; the AHA Go Red fund's first-ever investment. 15/ ECL - Yuval Bachar Led next-gen data center architecture at LinkedIn. Now builds off-grid hydrogen ones: 1,800 GPUs since 2024. 16/ Notejoy - Sachin Rekhi & Ada Chen Rekhi Sachin was Head of Product for Sales Navigator inside LinkedIn. Married co-founders, both ex-LinkedIn. 17/ KarmaCheck - Eric Ly He co-founded LinkedIn itself in 2002. Third act: background checks. $45M Series B on 600% revenue growth. Kafka, Pinot, DataHub, and Voldemort all ran inside LinkedIn before they were companies. Who else belongs on this list?
-
Mengjin Yan liked thisMengjin Yan liked thisOnce upon a time, all LinkedIn Espresso source was held in a single monolithic SVN repository and it was becoming very difficult to release new versions: Several months could pass between each release. Unhelpful graphs like this was used to tease it apart and split it into multiple repositories. This work was sometimes done after working hours in order to not disrupt anyone in the company because the main LinkedIn source "network" repository depended upon the Espresso client API. After many such sleepness nights, colleagues had to be informed the next day where to find the new repositories, if the migration and refactorings was successful. This change improved the team's agility by orders of magnitude and individual components could then iterate and release independently. It was probably much like changing every wheel of a car while it was being driven by other people. 🫠 In hindsight, was it worth it? For the team, definitely. For the individual, myself? Not so much. I probably should have asked for some recognition at the time.
-
Mengjin Yan reacted on thisMengjin Yan reacted on thisPerhaps one of the last 950 Maude LinkedIn Espresso team group photos in September 2019, before the COVID pandemic...
-
Mengjin Yan reacted on thisMengjin Yan reacted on thisLast week, I went to my new employer to start a new job. Looking back at the past several months, the job search was quite a journey. I applied to 95 positions, went through more interviews than I can count, and eventually received a few offers. I'm extremely grateful for the opportunities, the people who referred me, the recruiters and hiring managers who took the time to talk with me, and especially all the friends and former colleagues who encouraged me along the way. One thing I learned is that a job search is a very strange process. You can do everything right and still hear "no" many times. The number of rejection numbs you. And sometimes, the opportunity you really want simply doesn't work out because of timing, headcount, or circumstances completely outside your control. It is never my intention to share this just to celebrate a number, but to share a little encouragement with everyone who is still looking. I know many of my former colleagues and friends are still going through this process, and I know how difficult and exhausting it can be. Keep going. Keep learning. Keep reaching out to people. And don't let a rejection convince you that you are not good enough. Sometimes it really is just a matter of timing. I am rooting for you, all of you. I'm very fortunate that my journey has reached a good stopping point for now. For those still searching: hang in there. Your next opportunity may be closer than you think.
-
Mengjin Yan liked thisMengjin Yan liked thisEveryone knows that RDMA is fast. But how do you actually realize the speedups in your own application? As we were developing Ray Direct Transport (RDT), we found that there a lot of nuances in how Python applications use RDMA libraries like NVIDIA's NIXL, to the point where just 𝘂𝘀𝗶𝗻𝗴 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝘁 𝗹𝗶𝗯𝗿𝗮𝗿𝘆 𝗰𝗮𝗹𝗹𝘀 𝗰𝗮𝗻 𝗺𝗮𝗸𝗲 𝗮 7.5𝘅 𝗱𝗶𝗳𝗳𝗲𝗿𝗲𝗻𝗰𝗲 𝗶𝗻 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲. Our goal with RDT was to make RDMA easy and seamless to use for Ray applications like weight syncing for RL for LLMs. To do that, we also had to make sure performance was easy to achieve. Check out our blog post that walks through how to do this yourself using RDT. These steps are the key to the speedups that RDT achieves for weight syncing in end frameworks (𝘮𝘰𝘳𝘦 𝘰𝘯 𝘵𝘩𝘦𝘴𝘦 𝘪𝘯𝘵𝘦𝘨𝘳𝘢𝘵𝘪𝘰𝘯𝘴 𝘴𝘰𝘰𝘯): • 3.5s for Qwen3-235B-A22B with SkyRL - 18x faster than NCCL • 2.3s for GLM-4.5-Air with Miles - 2.2x faster than NCCL and 1.14x faster than Mooncake Blog link: https://lnkd.in/gCQSf-gz Joint work with my awesome teammates at Anyscale: Joshua Lee Xinyu Zhang Sumanth R Hegde Aaron Hao (and former teammate Dhyey Shah). Thanks also for feedback from Mengjin Yan Richard Liaw.
-
Mengjin Yan liked thisMengjin Yan liked thisCongratulations Philipp Moritz, Robert Nishihara.Edward Oakes, Richard Liaw, Ion Stoica and everyone else who made Anyscale what it is. You lost the A and the Y, but gained a lot more 😃. https://lnkd.in/gpg84b7bAnyscale signs definitive agreement to join NscaleAnyscale signs definitive agreement to join Nscale
Experience & Education
Volunteer Experience
-
Team Leader in Organising Conferences
International Aerial Robotics Competition
- 1 month
Science and Technology
Led a team organising the Opening Ceremony, Closing Ceremony and Technical Forum in the competition.
Courses
-
Advanced Cloud Computing
15-719
-
Cloud Computing
15-619
-
Computer Languages and Programming(91)
-
-
Computer Music
15-623
-
Computer Networks
15-641
-
Computer Networks
15-641
-
Computer Networks and Applications(92)
-
-
Computer Principles and Applications(91)
-
-
Data Structure(94)
-
-
Distributed Systems
15-440
-
Fundamental Pattern Recognition(94)
-
-
Fundamentals of Embedded Systems
18-342
-
Introduction to Computer Systems
15-513
-
Introduction to Machine Learning
10-601
-
Mobile and Pervasive Computing
15-821
-
Mobile and Pervasive Computing
15-821
-
Numerical Analysis and Algorithms(95)
-
-
Operating Systems(90)
-
-
Parallel Computing
15-618
-
Search Engines
11-642
-
Software Engineering(96)
-
-
Storage Systems
15-746
-
Storage Systems
15-746
Projects
-
Two‐phase Commit for Group Photo Collage
• Implemented the Server and the User Node of a multi-client group photo collage publication system using Two-phase Commit Protocol
• Made the system fault-tolerant to handle lost message and shutdown user node using timeout protocol and logging system
-
Scalable Service
• Designed and implemented a two-layer scalable web service with a database cache to start and shut down automatically according to the flow
-
File-Caching Proxy
• Designed and implemented the server and the caching proxy of a multi-client file system to operate multiple file-related C library calls
• Applied Session Consistency Model in the system as well as Least Recently Used (LRU) Caching protocol in the proxy
-
Transparent Remote File Operations
• Designed and implemented a multi-client Remote Procedure Call (RPC) system in C to handle multiple C library file operations remotely
-
Big Data Analysis Using Cloud Computing
Processed up to 67GB compressed Wikipedia request log data using AWS and MapReduce to find the trending topics
-
P2P Based Internet Chat System with a Central Server
-
Text-based Large Scale Search Engine
-
• Designed and implemented a text-based large scale search engine using Lucene API on corpus of 500,000+ documents from ClueWeb09 dataset
• Developed multiple query operators (NEAR, WINDOW, WAND, WSUM); implemented different retrieval models and algorithms including Ranked/Unranked Boolean, Okapi BM25, Indri, SDM, Pseudo Relevance Feedback, Letor; Applied several evaluation methods (MAP, P@n etc.)
-
Big Data Analysis Architecture and Analysis Model Based on Text Data
-
Independently proposed a big data analysis architecture and analysis model for text data, based on the general analysis framework for big data, text mining algorithms, researches on system architecture and the analysis of various cases.
Constructed and analysed the analysis model for a real case utilising the analysis architecture
-
Chatroom Discovery in NDN-based chat application ChronoChat
-
Systematically learned the basic architecture and security issues in Named Data Networking (NDN), ndn-cxx(NDN C++ library with eXperimental eXtensions) and the design of NDN-based chat application ChronoChat.
Independently designed and implemented the communication and storage protocol of the chatroom discovery feature in ChronoChat including the naming convention and type-length-value (TLV) package format of the data packets using gerrit for code review and github for version…Systematically learned the basic architecture and security issues in Named Data Networking (NDN), ndn-cxx(NDN C++ library with eXperimental eXtensions) and the design of NDN-based chat application ChronoChat.
Independently designed and implemented the communication and storage protocol of the chatroom discovery feature in ChronoChat including the naming convention and type-length-value (TLV) package format of the data packets using gerrit for code review and github for version control.
Wrote Unit Test test cases to test the chatroom discovery feature, integrated the feature into the whole application, and fixed some bugs originally in ChronoChat to improve performance.
Tested and presented the demo in NDNComm2014 (NDN Community Meeting). -
The Application of Intelligent Terminal in Energy Research
-
In this project, proposed and designed a specific structure of Smart Building System based on Cyber Physical System (CPS) which solved some of the challenge of CPS together with other member in the team.
Independently designed and implemented the storage layer using database normalization and MySQL.Other creators
Languages
-
English
Full professional proficiency
-
Chinese
Native or bilingual proficiency
Recommendations received
1 person has recommended Mengjin
Join now to viewOther similar profiles
Explore more posts
-
Sanchit Narula
The best advice I got on my promotions as a software engineer was from a Principal Architect at Amazon after he told me his story of getting promoted to Principal, and this has stayed with me my whole career. He said, very simply: > Your architecture can be a love letter to your technical brilliance, > but if it does not move customer and company goals, it will not move your level. That hurt a little. Because I recognised myself in it. A few years ago I wanted to design something with every shiny thing in the book: API Gateway, Lambda, DynamoDB, event bus, the works. On paper it was beautiful. In reality it was expensive, slow to ship, and did not match the skills of the team that would maintain it. So we shipped a hybrid architecture instead. Fewer moving parts, more reuse of existing services, cheaper to run, easier to hand over. Over time, here is how my thinking about promotions changed. 1. Start from the business goal, not the tech stack Before talking about Kafka vs SQS or MySQL vs DynamoDB, answer this first: What money, risk, or customer pain will this remove? Promotions track that, not the number of services in your diagram. 2. Ship finished products, not clever demos Leaders remember the feature that went live, stayed stable, and earned or saved real money. Drive the entire lifecycle yourself: design, build, launch, oncall, cleanup, docs. 3. Make the team faster, not just yourself Tools, scripts, templates, better CI, better dashboards. If people quietly say “things move faster when you are on this” your manager already has half a promotion case. 4. Do your share of the unsexy work Oncall, migrations, bugs, incident reviews. Senior folks who avoid this slow the entire org. Senior folks who lean into it become trusted very quickly. 5. Grow people, not just systems Mentor juniors, run design reviews, write clear docs, set standards. When your manager sees that you are already acting like the next level, formal promotion becomes a catch up, not a favour. 6. Be proactive about your promotion, not obsessed with it Have periodic one to ones. Talk about where you want to go and ask what evidence your manager would need to support that. Remember your manager has ten other people to manage. It is your responsibility to make your impact visible. Do not be desperate. Understand the timelines and work within them. As you get close to the boundary, collect the missing data points: impact metrics, incident write ups, design docs, peer feedback. Make the story easy to tell. Most importantly, work for the business, not for promotion-driven development. If your roadmap only exists to tick boxes on a promotion rubric, people can feel it. If your roadmap clearly creates value, saves cost, or reduces risk, promotion becomes the natural side effect. So instead of asking “What new tech can I use this year?” Ask, “Where can I create obvious business impact and leave the system and the people around me better than I found them?”
15 Comments
Explore top content on LinkedIn
Find curated posts and insights for relevant topics all in one place.
View top content