Today’s AI has a jagged intelligence surface. It is superhuman on well-defined benchmarks like the International Math Olympiad, but falls short in many real world domains.
Haize Labs brings superintelligence to the long tail of these domains through our agent foundry, which enables agent performance at parity with or better than domain experts.
Our agents are being deployed across many, many verticals. We are hiring an Applied Researcher to 1) turn companies into environments, 2) RL on these environments, and 3) rapidly scale and refine our products in existing verticals. This role spans research (RL, environment design, reward modeling), engineering (building, debugging, evaluating agents), and customer management (scoping, planning, relationship-building).
In this role, you will:- Push the pareto frontier of model capabilities
- Rapidly synthesize RL environments and reward functions / models
- Fine-tune agents to hillclimb almost any task in any business you can imagine
- Develop the research platform—observability, evaluation, custom reward models, simulation testing—connecting domain experts to models
- Love research, but also love engineering and making money
- Have a knack for probing model capabilities (e.g. benchmark papers)
- Want to bring intelligence to real people doing real work in the real world, not just Silicon Valley SaaS companies
- Have spent both time in a lab and in startup land
- Contribute to popular research frameworks (e.g. Prime-RL, Verl, Harbor, etc.)
- Spend their free time thinking about where AI is headed, and how to shape its capabilities
- Education requirement: doesn’t matter. If you are exceptional, you are exceptional.
- Location policy: NYC/SF.
- US visa sponsorship: If you are exceptional, we will sponsor.
$175,000 – $265,000 annual salary
And generous equity + benefits
What We Do
haize labs is the robustness and reliability layer underlying any AI model, in any industry, for any use case. to solve the robustness problem, we haize (i.e. stress-test and red-team) AI models to preemptively discover all their failure modes before production.
