Sourcerer

An AI technical-sourcing agent. It finds an engineer, researches them from public sources, and writes a grounded, cited fit-brief — every claim must point at real evidence, or it doesn't get made.

Pick a role to replay a cached sample run discover → research → synthesize, then read the brief — with the grounding score, clickable citations, and the “unverified” list of things it refused to assert. Source on GitHub ↗

The sample runs below use fictional personas (all links resolve to example.com) — the public demo never publishes a real person's data. To research a real, public GitHub candidate, run the tool from the CLI. If you believe you are shown here in error, contact us for removal.

How it’s built

Grounded, fail-closed. Every claim must cite a real evidence URL; anything the model can’t ground is demoted to the “unverified” list in code — not left to prompt discipline.
Clean async pipeline. discover → research → synthesize, with every external dependency (GitHub, web, LLM) behind a typing.Protocol and a deterministic mock — so the whole pipeline is unit-tested with zero network.
Safe by construction. Robots-aware, SSRF-guarded fetch (loopback / private / cloud-metadata blocked); public sources only; secrets from the environment.
Tested & shipped. A network-free test suite + CI, built test-first; GLM-5.1 via LiteLLM (swap the model with one env var).

A portfolio project by drinkerlabs.info · source & design docs ↗