TRM Labs · 2023 – 2026
Guardrails that make a whole team faster
Typed API contracts, internationalization checks and AI-agent tooling that caught problems before review, so a fast-moving team could keep shipping safely.
- Role
- Proposed and built
- Stack
- TypeScript, OpenAPI, GitHub Actions, Cursor, Claude Code
- merged PRs at TRM Labs
- 580+
The problem
TRM ships quickly, and AI-assisted development made it faster still: more PRs, from more people, touching more of the codebase. Speed is only an advantage if the mistakes it creates are caught early and cheaply.
Typed API contracts
Backend response shapes could change without a matching frontend change, and the mismatch surfaced in production. I wrote a script that generates types from our OpenAPI spec and runs in a pre-commit hook when backend code changes, with CI checks as well. A breaking change now fails type checks in the backend PR itself, where the backend engineer can coordinate the frontend fix before anything ships.
It is compile-time safety only; runtime validation and contract tests are the natural next step.
Internationalization that can’t drift
I led frontend internationalization. Because TRM’s users include governments, translations stay under the team’s control rather than raw AI output. Translations load lazily and are cached, keys are generated as TypeScript types so developers see real text on hover, and pre-commit and CI checks catch untranslated strings. Agent rules make AI-written PRs follow the same process.
Tooling for working with AI agents
- I split a 3,000-line top-level agent instruction file into focused skills that live next to the code they describe, and removed references to files that no longer existed.
- I wrote a skill for running independent tasks in parallel git worktrees, before cloud agents were common on the team.
- I wrote Cursor and Claude Code agents the team used day to day.
Testing culture
As PR volume grew, I pushed for stronger regression coverage, wrote unit and end-to-end tests, and proposed the workflow where PR checks flag regressions and deploy failures alert a channel. Another engineer built the end-to-end infrastructure; the shared goal was confidence at speed.
Impact
- Breaking API changes surface in the PR that causes them, not in production.
- Untranslated text cannot reach
main. - 580+ merged PRs at TRM Labs, shipped with these guardrails in place.