rodanaya/yangwenli
AI-powered corruption detection platform analyzing 3.1M Mexican federal procurement contracts (2002-2025). 16-feature per-sector calibrated risk model with 96% AUC, validated against 15 documented corruption cases worth 253B MXN.
What's novel
AI-powered corruption detection platform analyzing 3.1M Mexican federal procurement contracts (2002-2025). 16-feature per-sector calibrated risk model with 96% AUC, validated against 15 documented corruption cases worth 253B MXN.
Code Analysis
10 files read · 3 roundsA full-stack platform that analyzes 3.1M Mexican government procurement contracts using entity resolution (Spanish name normalization, phonetic matching, blocking), a multi-signal risk scoring pipeline (ARIA) that classifies vendors into 7 corruption typologies, and serves the results through a Fast
Strengths
The Hyperion entity resolution engine is genuinely well-crafted with domain-specific Spanish phonetic encoding, 80+ legal suffix variations, and a proper blocking strategy engine that reduces O(n²) comparisons. The ARIA pipeline is a real multi-signal scoring system with tiered investigation queues, ground truth validation against 748 documented cases, false-positive exception handling, and a full review workflow — not just a wrapper around a library.
Weaknesses
The core risk model (active_model.py, the actual scoring algorithm) was not visible in the explored files, making it hard to verify the claimed AUC-ROC scores and 'calibrated statistical model' claims. The analysis router is a 203KB monolith, and the 500+ _aria_cases_*.py scripts suggest the investigation case generation is largely manual/scripted rather than fully automated.
Score Breakdown
Signal breakdown
Innovation
Craft
Traction
Scope
Evidence
Commits
427
Contributors
1
Files
761
Active weeks
9
Repository
Language
Python
Stars
2
Forks
0
License
Apache-2.0