The Role
We're building production-grade agentic systems that audit medical claims end-to-end — reading raw medical records, reasoning over coding and clinical guidelines, and producing defensible findings that hold up to clinical and regulatory review. Reaching human-expert accuracy on noisy, long-context documents is one of the hardest unsolved problems in applied AI, and the field is moving weekly.
We're hiring an L4 AI Engineer who can step into an ambiguous problem, design an agent system from scratch, and ship it. You won't be plugging into someone else's architecture — you'll be deciding what the architecture should be.
What You'll Do
- Design agent systems from first principles. Decide the loop, the tools, the context strategy, the evaluation harness. Choose between single-agent and multi-agent topologies, between LLM reasoning and deterministic post-passes, between retrieval and direct context loading — and defend the choice with data.
- Engineer the context. The hardest part of building a good agent is what goes into the prompt and what comes out. You'll obsess over context windows, tool surfaces, structured outputs, citation grounding, and the prompt itself.
- Drive evaluation rigor. Build evals before you build the agent. Diagnose where it fails, fix the root cause, and prove the fix moved the metric.
- Use AI tooling like a power user. A meaningful fraction of your day will be spent driving Claude Code, Codex, and similar tools to plan, scaffold, refactor, and debug your own work. We expect you to be faster with these tools than most engineers are without them.
- Become a domain expert. Healthcare claims, coding guidelines, and the medical record itself are unavoidable parts of the job. Strong engineers who lean into the domain become outsized contributors here.
What We're Looking For
Required
- 2–4 years of applied ML / AI engineering experience with a Bachelor's in CS, Math, Engineering or equivalent — or a Master's in a similar program with no prior industry experience required. Either way, at least one production-quality system (industry, research, or substantial open-source) you owned end-to-end.
- Strong Python engineering. Clean abstractions, type discipline, async, tested code.
- Deep, hands-on understanding of agent loops — how a model decides to call a tool, how a tool result re-enters context, how loops terminate, where they fail.
- Hands-on experience with at least one major agent SDK — OpenAI Agents SDK, Anthropic SDK / claude-agent-sdk, LangGraph, or equivalent — and an opinion on the tradeoffs.
- Working knowledge of how modern coding agents are built and how they engineer context — what goes in the system prompt, how files are read and edited, how long-running tasks are planned and tracked, where they break.
- Fluency with Claude Code / Codex as a power user. You should be able to brainstorm, plan, and execute non-trivial engineering tasks with these tools — including reading their source when needed to understand or extend behavior.
- Solid command of VS Code and git — branches, rebases, worktrees, conflict resolution, PR workflows. Not optional.
- A bias toward measurement: you don't ship without an eval, and you don't believe a number you can't reproduce.
Strongly preferred
- Experience designing structured outputs (Pydantic / JSON Schema) and tool interfaces that LLMs reliably call correctly.
- Familiarity with reasoning models (o-series, Claude extended thinking, Gemini thinking) and a sense of when they earn their cost.
- Prior work on long-context, citation-grounded systems where the model must point to evidence, not just answer.
- Healthcare, legal, finance, or any other domain where "mostly right" is unacceptable.
Nice to have
- Document understanding (OCR, layout-aware models, table extraction).
- Vision-language models, multimodal retrieval.
- Production experience with caching, observability, and cost control on LLM workloads.