Build the physical intelligence layer for the industrial economy.

Our vision systems run at steel mills, scrap yards, and recycling plants. Behind them, the agent trains and evaluates our models, finds their weaknesses, and searches millions of images for examples of rare classes.

It is already our main way to investigate and iterate on our deployed models and our research. It works, but it can become a much more powerful tool.

This is not a prompt-engineering role. Most of the work is systems and measurement: deciding what the agent can see and do, and proving on a replayed set of real questions that a change improved its answers.

Your teammates already contribute comfortably in a mix of programming languages, and we expect you to do the same. You'll work in person about 3 days a week at our office near Union Square, NYC.

• Post-train open-weight models on our own tasks: SFT on curated traces, then RL with verifiable rewards (GRPO and its variants, and DPO-style preference methods where they fit). Build the reward functions, the rollout environment, and the eval gate, then serve the result.
• Own the agent harness: context management over long multi-turn sessions, tool-call budgets, background tools, failure recovery, and the policy that decides which tool calls need approval.
• Grow the tool catalog (SQL over Postgres and ClickHouse, frame retrieval, segmentation, depth, model training and evaluation), and give each tool a typed schema, clear errors, and a test.
• Run the evaluation loop: replay real questions against a candidate harness or model, grade the answers with calibrated judges, and report quality and dollars per correct answer before anything ships.
• Push open-source models beyond their weight class with routing, test-time selection, and advisor and planner/worker setups in which a strong model plans and cheap ones execute.

• You've shipped a tool-calling LLM or VLM agent to real users, and you can explain how you evaluated it and how it failed.
• You've post-trained open-weight LLMs or VLMs yourself with SFT and at least one RL method (GRPO, PPO, DPO, or similar), and you can explain what your reward or data changes did to the eval.
• You understand computer vision fundamentals: detection and segmentation metrics, and the common architectures.