Job at a glance
Job Description What you'll do Own end-to-end slices of the agent: prompt and context architecture, retrieval over policy and coverage data, tool and API calls into core claims systems, and handoff to a human when the agent should stop. Build and run the eval loop, golden sets, trajectory evaluation, LLM-as-judge, error analysis on real production call traces, and make it the thing that gates releases.
Drive containment and resolution rates up without trading away accuracy, and instrument the system well enough that we can tell which of the two moved. Work directly on the voice path: turn-taking, interruption, latency budgets, and the telephony layer underneath (LiveKit, SIP). Sit in on carrier calls and listen to recordings. The best fixes usually come from hearing where the conversation went sideways, not from reading the logs.
What we're looking for Building and deploying AI applications. You've shipped an LLM application that real users depend on, and you know that the hard part is that the output is not predictable. You're fluent in the building blocks LLMs, context engineering, RAG, agentic workflows, and you have real ML and deep learning depth underneath that fluency. Most importantly, you know how to make an unpredictable system behave predictably enough to trust: disciplined evals, error analysis, and statistical measurement rather than vibes and a handful of spot checks.
Software engineering fundamentals. You can reason about cost, latency, reliability, and scalability as tradeoffs rather than goals, and you know which ones you're making. You've designed data stores and distributed systems, you write tests, and you take security and privacy seriously while advocating CI/CD process automation. We handle claims data, recorded calls, and PII under carrier and regulatory scrutiny.
Using coding agents. You work with coding agents daily and you're good at it: managing their context, deciding when a spec is worth writing and when it isn't, giving them verifiers so they can close their own loops, and knowing when to intervene versus leave them alone. You also keep changing how you work as the tooling changes, rather than settling into a 2025 workflow. Shaping the build. You won't be handed a finished spec.
You'll be in the room where we decide what the agent should do next, which means you need product sense and enough insurance context to argue about it. You'll know when to put an MVP in front of a carrier next week and when a claims workflow demands that we slow down and get it right the first time. Underneath all four: you keep learning. This field looks different every six months and we expect to keep changing our minds.
Nice to have Real-time or streaming voice systems (ASR, TTS, telephony, SBC/SIPREC) Insurance, claims, or another regulated domain where being wrong has a cost attached Experience building internal eval or observability tooling other engineers actually adopted. Experience in Machine learning principles, or have experience in model experimentations, training and deploying models in Productions. Experience in P&C Insurance or exposure to Guidewire products and services.
Prompt or pipeline optimization frameworks (DSPy, GEPA) applied to a production system How to apply Send a short note about something you shipped where the model's unpredictability was the central problem, and what you did about it. We care more about that than about a list of frameworks.