About Midpage
Midpage is the search engine for legal data used by AI labs. We cover all US court data - 20M records. Over 300 law firms use our platform directly, 200k+ visitors read cases on our site every month, and five multibillion-dollar companies including Perplexity trust us as their legal data supplier. We're a team of 7 in Bowery, lower Manhattan. Our ARR has grown from $400k to $2M in the last 4 months.
The role
We're seeking an engineering generalist to build the first RL environments and benchmarks purpose-built for long-horizon legal reasoning—tasks where AI agents must search, read, analyze, and draft across real case filings, the same work that still takes teams of lawyers days to weeks. Frontier labs are will use these environments to make future models more legally capable and we need an engineer to own the infrastructure that makes it all work.
You'll design and scale the systems that turn millions of real court filings into verifiable evaluation environments and RL training tasks. You'll work directly with our attorneys, our data pipeline, and our partners at frontier AI labs. Your job is not mainly about RL training itself, and more about building a RL environments factory that uses our data to generate difficult RL environments.
What you'll do
- Build and maintain the evaluation harness and RL environment infrastructure—task runners, sandboxed environments, and scoring logic that can scale to thousands of parallel agents.
- Own the data pipeline that turns freshly collected court filings into benchmark and RL tasks before they reach any model's training set.
- Integrate with partner harnesses and model APIs to run contamination-free evaluations.
- Collaborate with attorneys to translate legal workflows like cite checks, motion drafting, and precedent research into structured, scorable task formats using the Harbor spec.
What we're looking for
- Strong generalist software engineering fundamentals. You've built, scaled, and maintained diverse systems in production.
- You've built entire systems yourself, don't require detailed specs or product managers, and take full ownership over your projects.
- Deep experience with Python, bonus for TypeScript. Most importantly, you can work on hard engineering problems.
- You should be kind, self-managing, and a clear communicator.
- You make effective use of Cursor/Claude Code/Codex and are capable of writing good code without them.
Bonuses but not requirements
- Familiarity with LLM evaluation. You get what makes a good rubric and why benchmarks leak.
- Comfort working with messy, real-world document data (legal filings, PDFs, long-form text).