About Us

Fieldguide is establishing a new state of trust for global commerce and capital markets by automating and streamlining the work of assurance and audit practitioners — specifically in cybersecurity, privacy, and financial audits. We build software for the people who enable trust between businesses.

We're based in San Francisco, CA, and we're backed by Goldman Sachs Alternatives, Bessemer Venture Partners, 8VC, Floodgate, Y Combinator, and more. Over 50 of the top 100 accounting and consulting firms trust Fieldguide to power mission-critical work.

About the Role

You'll join a genuine 0→1 team on the ground floor of one of the company's biggest new bets. This seat is specifically product-focused: you'll own agent quality, ship agents that do real audit work, and work alongside practitioners.

Depending on your experience and what you're looking to own, you may join building within a major agent area, owning one end-to-end, or setting technical direction for agentic audit work across the team. We're hiring across all levels and will calibrate during interviews based on scope and demonstrated experience.

What You'll Do

• Make agent judgment repeatable: run error analysis on real testing data and turn findings into concrete fixes

• Tradeoffs such as quality/latency/cost across a long multi-phase run

• Build structured-output pipelines that turn model output into real audit artifacts

• Take ambiguous problem statements and turn them into a plan, a shipped feature, and a clear read on what was cut and why

• Work directly with an embedded subject matter expert and with design-partner firms, turning their feedback into agent changes within days

• Expand agent coverage into new controls and new areas of internal audit

Who You Are (All Levels)

• Product-minded and full-stack: you've shipped LLM-backed features to production against real users, and you measure yourself on whether they got used

• You're fluent in evals and error analysis, and you apply them in service of shipping something practitioners trust

• You have real opinions on model selection, prompting, and orchestration tradeoffs, and you can defend them with evidence rather than vibes

• Energized by 0→1 work: you'd rather define the problem than inherit a spec, and you don't stall on ambiguity

• Strong instincts for human-in-the-loop design