AI-Powered Compliance Monitoring — UK FCA-Regulated FinTech

ML Transaction Scoring Engine

Python scikit-learn gradient boosting model trained on 18 months of historical transaction data (1.2M labelled transactions — compliant, suspicious, confirmed fraudulent).

Features

• transaction amount, merchant category, geographic pattern, velocity (transactions per hour), peer comparison (deviation from cohort behaviour), device fingerprint, and time-of-day pattern.
• Model deployed as FastAPI microservice.

FCA JMLSG Alignment

Three-tier risk categorisation: Low (automated pass), Medium (enhanced monitoring — automated + analyst review queue), High (SAR candidate — mandatory analyst review before any automated action).

JMLSG guidance

• no fully automated adverse decisions on customers without human review.
• All High-risk flags routed to human analyst within 4 hours.

OFSI/HMT Sanctions and PEP Screening

• Real-time sanctions screening: OFSI UK (Office of Financial Sanctions Implementation) consolidated sanctions list, HM Treasury UK sanctions lists, and Dow Jones World-Check for PEP (Politically Exposed Persons) screening.
• Fuzzy matching algorithm (Levenshtein distance + phonetic matching) to catch name variations and transliterations.

Every transaction score logged

• model version, feature values at time of scoring, score, risk tier, human review outcome (if any), and SAR reference (if generated).
• Immutable audit log (AWS CloudTrail + S3 Object Lock) — the FCA can reconstruct any decision at any point in time.