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.
About You
This role is a blend of two traits that are critical for success. Our system has to be genuinely novel and boringly reliable at the same time, and an engineer with only one of those instincts pulls the pod off balance.
You are AI-native and inventive:
- Frontier models are instruments you play daily — in how you build (coding agents, multi-model workflows, letting an agent draft while you direct) and in what you build (planners, classifiers, tool-calling systems).
- You come at problems with approaches nobody asked for, and you are willing to be wrong out loud in service of moving the work.
- "The model can probably do this" is a hypothesis you test with an eval, not a hope you ship on vibes.
- You prototype fast, generalize what survives contact with reality, and delete what doesn't.
You are steady in delivery and reasoning:
- You ship in small, traceable increments, week after week; teammates can find the ticket from your branch name and the reasoning in your design doc.
- You write your thinking down — design docs before lynchpin systems, and review comments that catch what tests miss: the uncalibrated confidence score someone will read as a guarantee, the trust boundary conflated with a durability guarantee.
- Production is yours. You fix the OOM at the right layer and treat a correctness rework as finishing the job, not a "fast follow".
- You simplify your own work: deleting your days-old code because a simpler approach developed is a win, not a loss.
Table-stakes:
- Production service fundamentals: API design, data contracts, authorization boundaries, observability.
- Hands-on experience with LLM agent systems — tool-calling patterns, MCP, the Anthropic SDK, or equivalents — running in front of real users.
- Fluency in a strictly-typed codebase.
- You put safety properties in code, not in prompts — and you can say why.
- Clear written communication about tradeoffs; here, decisions live in documents and threads.
- Prior experience in and passion for early-stage startups and/or high-growth environments.
Nice-to-haves:
- Experience with durable-execution engines in production.
- Event-driven systems with schema governance — event bus patterns, pub/sub, schema registry, Avro/Protobuf.
- Eval frameworks and LLM observability.
- Building and consuming MCP servers.
- Data lake or warehouse-adjacent data engineering.
- A regulated domain — insurance, fintech, healthcare — where correctness is contractual.
- Frontend experience; it's where our users live.
Vouch provides several benefits to help you bring your best self to work:
- 💰 Competitive compensation and equity packages
- ⚕️ Health, dental, and vision insurance
- 🍼 Parental leave
- 🌴 Flexible vacation time
- 🪷 Wellness allowance
- 🛜 Technology allowance
- 📚 Company-sponsored personal and professional development
- 🏫 L&D: Partnerships with Ethena and monthly Lunch & Learns
- 🧘 Wellbeing: access to many wellbeing perks, including Peloton, Fetch, OneMedical, Headspace care+, etc.
- 🤗 Caregiver Support: company seed into the dependent care FSA and company sponsored Care.com membership.
- 📊 Regular performance reviews: Vouch conducts regular performance discussions with all team members, offering goal setting and check-ins, development discussions, and promotion opportunities.
What to expect in a typical interview process:
(Please note these steps may vary slightly depending on the role)
- 30-minute phone call with our recruiting team
- 30-45 minute video interview with the hiring manager
- Case study/technical screen
- Meet the team! 30-45 min 1:1 video discussion with 3-4 team members you’d work closely with in the role
- Executive chat
Compensation philosophy:
Our salary ranges are based on paying competitively for our size and industry and are part of our total compensation package, which also includes benefits and other perks. We also include stock options in all compensation packages and believe all Vouch employees should have the opportunity to become owners in the company. Individual pay decisions are based on a number of factors, including qualifications for the role, experience level, skill set, location, and business need. The pay range provided is subject to change and may be modified in the future.
Vouch believes in putting our people first, and building a diverse team is at the front of everything we do. We welcome people from different backgrounds, experiences, perspectives, and ranges of abilities. We are an equal-opportunity employer and celebrate the diversity of our growing team.
If you require reasonable accommodation to complete this application, interview, complete any pre-employment testing, or otherwise participate in the employee selection process, please direct your inquiries to [email protected].