Founding Engineer — ML Platforms Engineer
About the role
ML Platforms Engineer role at Cumulus Labs focusing on building and running the GPU orchestration layer for production AI, including workload scheduling, GPU allocation, and multi-cloud fleet management. Involves end-to-end ownership, hands-on system design, and in-person collaboration with a small founding team in San Francisco.
Responsibilities
- Build and extend our GPU orchestrator: scheduling, fractional allocation, live workload migration across GPUs with no downtime
- Design and evolve multi-tenant primitives: quotas, isolation, usage metering, a tenant-facing inference gateway
- Own observability for the fleet: metrics, logs, and traces at scale
- Debug hard, systems-level problems across the stack, from scheduling logic down to GPU memory and networking
- Make real architectural decisions, not just implement someone else's design
- Ship fast, own your systems end to end, and work directly with the founder
Qualifications
- Excellent fundamentals: data structures, algorithms, distributed systems concepts, and the judgment to apply the right pattern to the right problem
- Real production experience, ideally with systems that had to stay up and scale under load
- Strong design instincts: you can reason about tradeoffs, not just follow a framework's conventions
- Fast learner who can go deep in unfamiliar territory; specific experience with Go or Kubernetes is a plus, not a requirement
- Comfortable using modern AI coding tools (Claude Code heavily) to move fast without losing rigor
- Willingness to work in person in a small team, solving problems nobody has solved before
Skills
About Cumulus Labs
The Fastest Multimodal Inference OS
Cumulus Labs is a fast multimodal inference provider, purpose-built for AI teams who want faster performance, lower costs, and zero infrastructure work on fine-tuned & open source models. Most teams today are stuck choosing between bad options. Self-hosting inference means wrestling with configurations and babysitting infrastructure that slows/breaks at scale. Big providers like Fireworks are convenient but extremely expensive and idle GPUs. Cumulus ships Ion, a proprietary inference engine that run LLMs, VLMs, and audio/video gen with high performance and lower cost.