AudaciousMeristem, a new frontier AI lab.
Meristem is a new frontier AI lab founded by ex-DeepMind researchers Kelvin Guu and Kenton Lee, backed by Audacious and Lightspeed. They need someone to own the platform that lets the research move fast.
Kelvin Guu
Google DeepMind researcher and Stanford PhD; foundational work on retrieval-augmented language models (REALM).
Kenton Lee
Gemini research scientist; led RL for computer use; core image-understanding & pre-training work; co-authored ELMo and contributed to BERT.
Own the foundation the whole lab runs on.
This is a frontier AI lab looking for a DevOps & Platform Infrastructure Engineer who thrives in the complexity of early-stage systems and brings both craft and pragmatism to every layer of the stack. Right now the infrastructure is patched together well enough to shield the research and product teams from pain, but they know where it’ll break as they keep building.
You’ll own the foundation that lets the rest of the team move fast, stay productive, and stay secure: developer tooling, GPU workload orchestration, secrets management, observability, and the platform reliability that makes running many concurrent AI experiments feel routine. They want a proactive entropy-reducer, someone who spots the weak joints before they fail, fixes them autonomously, and raises the ceiling on what the whole team can do.
The whole platform layer.
Developer productivity
Reduce friction for AI/ML researchers; turn manual deployment steps into repeatable, observable workflows anyone on the team can use and trust.
Platform reliability & observability
Set up alerting, monitoring, and on-call so issues surface before outages; tune alerts and own the feedback loop between signal and fix.
GPU compute infrastructure
Orchestrate GPU resources across providers (Modal, Baseten, neoclouds); support experiment sweeps and, eventually, production training runs.
Cloud & cost hygiene
Manage workloads across AWS and GCP; proactively spot waste and right-size resources as the lab grows.
Security
Own networking, secrets management, and least-privilege access; harden the boundary between internal services and external compute.
A 0→1 builder who reduces entropy.
- You’ve built infrastructure 0→1, set things up correctly the first time, not inherited a mature platform.
- Autonomous and proactive; you don’t wait to be scoped, you find the entropy, write the task yourself, and close it.
- Strong systems intuition: you can look at a patchy set of services and reason clearly about where it breaks under load.
- Genuinely curious about ML systems; you collaborate closely with researchers on experiment infra without needing everything explained.
- Comfortable being the first or only platform person for a meaningful stretch, energised, not intimidated, by that.
- You communicate clearly about risk and trade-offs, the team is never surprised by infra debt they didn’t know was accruing.
Deep in several; comfortable picking up the rest.
AWS / GCP · serverless compute · neoclouds (GPUs)
Secrets management · password manager · software supply-chain security
VPN · MDM
Kubernetes / Slurm · Terraform / IaC
Nice to have: ML training infra (experiment management, distributed training, checkpoint storage) · high-reliability / high-throughput backgrounds (fintech, large-scale web) · evaluating and migrating between compute platforms · familiarity with AI agent systems, inference serving, or LLM deployment.
This role is probably not for you if…
- You’re primarily motivated by fundamental research, this is DevOps / infrastructure, not a path to model-training algorithms.
- You want to work on product features or business logic, the scope here is the platform layer.
- You need a well-scoped backlog, the projects here are often ones you’ll identify and define yourself.
- You want infrastructure that’s already mature, it isn’t, and there will be stretches where things are messy before they’re clean.
- You want to specialise deeply in one narrow system, this role spans security, compute, observability, and developer tooling.
Who’s behind it, and how to get in.
Meristem, founded by ex-DeepMind researchers Kelvin Guu and Kenton Lee.
Interview process: 1–2 intro calls (mutual fit, background, role context) → 45-min system debug → on-site deep-dive (system design, system debug, and team intro).
Own the platform a frontier lab is built on.
San Francisco, in office. The company is in stealth, reach out for an intro and we’ll share more under wraps.