CASE STUDY / FINANCIAL AI
AegisAgent
Fraud Investigation & Evidence Review
Fraud alerts with evidence and human review.
Research project01 / The problem
A model alert needs evidence and accountable review; it cannot itself establish regulatory suspicion.
02 / My approach
A Python pipeline scores synthetic transactions, while a Streamlit interface and FastAPI case service support review. Narrative drafting uses AWS Bedrock with evidence and privacy controls.
- Transaction features
- Fraud scoring
- Evidence & review
- Grounded draft
03 / Engineering decisions
- Keep development and locked evaluation windows distinct.
- Preserve human roles, rationale, case history, and evidence integrity.
- Fail closed at the production identity boundary rather than treating local identity as production authentication.
04 / Results & current state
Locally runnable portfolio/research project; current documentation says it is not publicly deployed.
Phase 1 locked-tail evaluation
Run date unavailable. Report has no run timestamp; model version v_20260828_002211. Model card last updated August 29, 2026; inspected September 8, 2026.
138,930 synthetic-data rows; prospectively locked tail, not historically pristine or an external test. Metrics are not banking production performance.
Original locked evaluation report05 / Limits of the evidence
Not a Node.js application, production banking platform, AML decision engine, or FINTRAC-compliant filing system. Model alerts are not human determinations of regulatory suspicion. Evaluation uses public synthetic transactions. The locked tail is not historically pristine because earlier revisions reported aggregate metrics over the full test file. Model 4 outperformed the stacked ensemble on locked-tail PR AUC; neither result establishes external banking performance.
Public sources
Reviewed September 8, 2026. Links are pinned to the inspected public repository revision.
Current project documentationModel card and scopeFastAPI service architecture