AI Engineer

The Role

Vouch is building AI software for judgment-heavy insurance work: a system that learns from experts and gets measurably better week over week. We are early — a small team, real experts, real production usage, real customers, and a lot of unanswered technical questions.

This is genuinely interesting work, and we say that well-aware of how every job posting claims that. We are betting on a specific approach to AI in a regulated domain, and we are not going to lay it all out in this posting — we'd rather it stay our edge until you're across the table from us. The hard problems live exactly where you'd hope — making an LLM system durable, auditable, and measurably improving, in a domain where being wrong has consequences.

You would join early in the system's life, in a rapidly evolving codebase that already carries more test code than source code. That ratio is not an accident; it is our style.

How we work, concretely: reasoning is written down and public — design docs land as pull requests, root-cause writeups happen in the channel, and demos are async videos every Friday morning. We ship to staging many times a day and to production behind consent-based pushes; standups are short and bot-recapped, and the real arguments happen in threads and design-doc reviews.

What You'll Do

• Ship agentic workflows end to end. Design tool contracts, capability boundaries, and approval gates for an agent doing real insurance work under human judgment — then carry your change through review, deploy, and production ownership.
• Build on a durable-execution backbone. Our workflows run with pinned worker versioning and a replay-compatibility gate in CI; a crashed worker has to resume mid-workflow without losing state. You'll extend that substrate and understand it deeply.
• Make model behavior measurable. Deterministic corpus tests are the merge gate; eval suites are the diagnostics; characterization corpora pin behavior before refactors. You'll maintain and grow that machinery, along with LLM tracing and token/cost observability.
• Treat prompts and tool definitions as engineered artifacts — versioned, cache-stable, snapshot-tested, and reviewed like code, because they are code.
• Run the event and data substrate. Event bus with schema-governed domain events, object storage, OLTP/OLAP databases, and the two-way plumbing between an agent and the systems of record it must respect.
• Work with AI, on AI. This team builds with frontier coding agents as daily instruments, and our specs are written to be read by humans and coding agents alike. You'll do both: use the tools hard, and build the system that makes an AI coworker trustworthy.
• Review with teeth. Our review culture prizes finding the silent failure path — the empty string that detonates three stages later — before production does.