What it is
A careful, reproducible look at ranking with a decision model.
What it does: Scores each posting against every candidate in one call and checks the ranking against an answer key.
Why it belongs in the showcase: Clear numbers, cheap to rerun, and honest failure analysis.
Caveats: Not for real hiring decisions; AI-labelled key.
Evidence
Public repo created 2026-10-19; code is MIT, data has a separate licence file. README reports 400 postings (0 failed), about 197 s, 3,988,299 input tokens, about US$0.17, top 3 vs answer key 96/100, top 1 with the 1.5 cutoff 45/100, and "none" detected 14/14. The answer key was labelled by Claude and spot-checked by the author. 16 of 26 top-1 misses were seniority mix-ups. `python avaliar.py` recomputes the numbers without API calls.
Caveats
- Experiment only, not a hiring tool; keep people in the hiring decision
- Answer key labelled by an AI model and only spot-checked
- All candidates are fictional; file names and some notes are in Portuguese
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.