Featured project
Outbound
Audit coding-agent transcripts for sensitive information sent to model providers.
Shipped CLIAn open-source local command-line tool, with detection metrics and regression tests documented in the repository.
My role: Solo
The problem
Coding-agent conversations can include sensitive file contents and command output. Session transcripts hold the record, but reviewing them by hand is difficult.
What I built
- Built a local CLI that scans agent transcripts for sensitive information sent to model providers.
- Reports what was exposed, when, and whether it came from a file read, a command, or a pasted prompt.
Engineering decisions
- Keep raw findings out of the database; store redacted excerpts and keyed fingerprints instead.
- Use Node’s built-in SQLite support with no runtime dependencies.
Verification
- A synthetic labeled corpus gates CI on per-category precision and recall.
- False-positive cases found in 1.7 GB of real transcripts became regression tests.
- A test reads the database bytes to check that raw sensitive values were not persisted.
Tech stack
- TypeScript
- Node 22
- node:sqlite
- Vitest
- Zero runtime deps
- Stage
- Shipped CLI
- Role
- Solo
- Platform
- Local CLI
- Year
- 2026
