Principal Engineer

Why This Role Matters:

Vouch's engineering organization is becoming agent-native: a large and growing share of our production code is written by AI coding agents working inside human-governed guardrails. That moves the hard problems. Writing code got cheap; deciding what to build, keeping the architecture coherent, and proving that what shipped is correct did not. This role owns that layer.

This is a principal-level individual contributor seat reporting directly to the CTO, with a mandate that spans our estate: a harness for agent-built software — your first ownership deliverable — production AI doing judgment-heavy insurance work, a backend whose bounded contexts need to become real, and the cloud platform underneath all of it.

We are not going to lay the whole system out in a posting — we would rather keep the details as our edge until you are across the table from us. The short version: we extracted the machinery that let a small team ship roughly 1,100 governed changes in seven weeks at a sub-1% change-failure rate, and we are generalizing it into a platform – leaving open technical questions are the kind this role would actually want to be responsible for answering.

• Own the platform that governs agent-built software. A harness where markdown specifications compile to a traceability lock, deterministic gates decide "done," every agent action leaves tamper-evident evidence, and a calibrated multi-vendor panel of LLM reviewers blocks what humans should not have to catch. You will take it from working extraction to adopted platform: scope its next milestone, build its plugin surface, operate and score the review seats, and prove it generalizes beyond the repository it was born in.
• Make bounded contexts real across our backend. Our backend estate asserts "one service = one bounded context"; a handful of services actually implement ports and adapters; the rest is your brownfield. You will turn design-doc DDD into enforced architecture — context maps, anti-corruption layers, boundary checks that run in CI — without freezing a team that ships daily.
• Ship agentic systems that do real insurance work. Durable multi-week agent workflows with human approval gates, tool contracts and capability boundaries, planner/executor orchestration — extended in production, in front of the experts whose judgment they encode.
• Make model behavior measurable — and autonomy earned. Eval suites as merge gates, golden datasets, LLM tracing and token/cost observability, calibration of LLM reviewers against ground truth. Autonomy expands when the evidence says so, and you own the evidence.
• Run platform and delivery engineering at organization scale. An AWS multi-account estate in Terraform behind Git Ops; two orchestrators in production; CI/CD pipelines that are generated, not hand-written; ephemereal environments per pull request. You will make the paved road faster than the shortcut.
• Be the CTO's architecture thought partner — and raise the bar. Strategic-design working sessions, ADRs with teeth, review comments that catch what tests cannot; specifications written to be read by humans and coding agents alike, mutation-testing bars, adversarial review. When we argue about where a boundary belongs, you are the other half of the argument.

A principal engineer here is equal parts AI engineer, Dev Ops engineer, and software architect. We mean that literally, and we will interview for all three.

• You have shipped domain-driven design, not just read about it: a bounded-context map you drew and then enforced, an anti-corruption layer that saved a migration, a hexagonal refactor that survived contact with a deadline — and you know when the pattern is overkill.
• You are at home in more than one language ecosystem — e.g., systems-level work in Rust or a peer language, large-scale strictly-typed Type Script, etc. — and you know which discipline each one demands.

• You have built and operated agent systems in front of real users — tool contracts, capability boundaries, approval gates — and the machinery that makes them trustworthy: eval gates, provenance capture,…