Software Engineer, Model Serving Infrastructure

Software Engineer, Model Serving Infrastructure About Anyscale: At Anyscale https://www.anyscale.com/, we're on a mission to democratize distributed computing and make it accessible to software developers of all skill levels. We’re commercializing Ray https://docs.ray.io/en/latest/, a popular open-source project that's creating an ecosystem of libraries for scalable machine learning. Companies like OpenAI https://thenewstack.io/how-ray-a-distributed-ai-framework-helps-power-chatgpt/, Uber https://www.uber.com/blog/horovod-ray/, Spotify https://engineering.atspotify.com/2023/02/unleashing-ml-innovation-at-spotify-with-ray/, Instacart https://www.youtube.com/watch?v=3t26ucTy0Rs&list=PLzTswPQNepXmLUiL4F_1VHrPcCz1OeILw&index=23&pp=iAQB, Cruise https://www.youtube.com/watch?v=gj0BqvfX_wI&list=PLzTswPQNepXmLUiL4F_1VHrPcCz1OeILw&index=46&pp=iAQB, and many more, have Ray in their tech stacks to accelerate the progress of AI applications out into the real world. With Anyscale, we’re building the best place to run Ray, so that any developer or data scientist can scale an ML application from their laptop to the cluster without needing to be a distributed systems expert. Proud to be backed by Andreessen Horowitz, NEA, and Addition https://www.wsj.com/articles/ai-startup-anyscale-adds-99-million-to-andressen-horowitz-led-funding-round-11661254200 with $250+ million raised to date. About the role: Anyscale is actively seeking talented engineers to join our team and contribute to the development of next-generation, high-performance machine learning serving systems. We value diversity and inclusion, and we encourage individuals from underrepresented groups to apply. Many existing ML serving tools are inherited from previous infrastructure generations, but emerging ML applications present new requirements, such as high compute demands, specialized hardware needs, and the integration of multiple models and business logic within a single request. At Anyscale, our mission is to provide a powerful yet simple set of tools that enable the seamless deployment of complex ML applications in production. The Challenge: What if you could build the infrastructure that powers AI applications for millions of users worldwide? Ray Serve is the production-grade serving framework that makes this possible—and we need exceptional engineers to push its boundaries. You'll be working on problems that sit at the intersection of distributed systems, machine learning, and high-performance computing. This isn't about maintaining CRUD apps or tweaking configurations—this is about solving fundamental computer science problems that directly impact how the world deploys AI. What You'll Actually Build: Example projects: - Asynchronous inference: Let the client submit a request and get a request handle that asks for its requests completion while not blocking the client side. Really important for image, video, or audio generation applications. - Sub-millisecond Model Routing: Design and implement intelligent request routing systems that dynamically balance load across thousands of model replicas while maintaining strict latency SLAs - Zero-Downtime Model Updates: Build sophisticated traffic management systems that seamlessly transition between model versions at scale, handling terabytes of inference requests without dropping a single query - State Management at Scale: With many models and many replicas deployed into production, the control loop’s state management can become the bottleneck for events such as routing, autoscaling, etc. What are the architectural improvements that can shift the envelop of scale by 10x going from 1000s replicas to 10,000s replicas, etc. - Multi-Model Orchestration: Architect frameworks for complex ML pipelines where dozens of models need to communicate, share resources, and maintain end-to-end latency guarantees - Observability & Debugging: Build deep introspection tools that make it trivial to debug distributed ML applications—be ... (truncated, view full listing at source)