Case Study
WalkerFinance
RAG + ML finance system — automation over personal financial data that only acts when provably better than existing rules.
Metrics
- 2-gate
- ML classifier
- 50+
- human labels to activate
- 0.70
- confidence floor
Problem
Trustworthy automation over personal financial data — ML that earns its place.
Stack
- Multilingual RAG (sentence-transformers to pgvector)
- Two-gate scikit-learn classifier (overrides rule engine only after 50+ human labels, beating baseline on 20% holdout, 0.70 confidence floor)
- 5%-contamination IsolationForest
- 95%-confidence OCR
- Zero-trust: JWT, mTLS, default-deny RLS
Constraints
ML must be provably better than the existing rule engine before any override.
No silent failures.
Outcome
ML that only acts when it is provably better than rules.
Want something like this?
Start a project