BFSI ML Systems — Transaction Intelligence, Underwriting & Fraud

Overview

Designed and shipped a portfolio of production ML systems for enterprise banking clients across UK and Indian markets — covering transaction intelligence, fraud detection, underwriting automation, and next-best-action recommendations. All systems run against live transaction volume with strict compliance, explainability, and privacy constraints.

Key Systems

• Transaction Categorization Engine — 92% classification accuracy across a 3-level hierarchy, running over 500K+ monthly transactions. Includes CO₂-impact tagging for sustainability reporting in the PFM product line.
• Next Best Action (NBA) — recommendation system for retail lending that drove a 25% uplift in loan-product uptake among targeted customer segments.
• Underwriting Acceleration — ML-assisted risk-scoring models that cut underwriting review time by 40%, while preserving lender-required auditability and explainability.
• Fraud Detection — ensemble model replacing rule-only screening; delivered a 30% reduction in false positives without degrading fraud-catch rate.
• Privacy-Preserving GenAI Chatbots — pseudo-anonymization pipelines paired with Small Language Models (SLMs) so customer-facing conversational AI could run with on-prem data privacy guarantees.

Architecture & Approach

flowchart LR
    TX["Live Transactions<br/>500K+/month"] --> ING["Ingest<br/>pseudo-anonymize PII"]
    ING --> FE["Feature Pipeline<br/>Apache Airflow"]
    FE --> MODELS["ML Models"]
    MODELS --> M1["Categorization · 92%"]
    MODELS --> M2["Fraud · -30% FP"]
    MODELS --> M3["Underwriting · -40% time"]
    MODELS --> M4["Next Best Action · +25%"]
    MODELS --> MON["Monitoring + Guardrails<br/>drift · fallback · MLflow"]

• Python + FastAPI services backed by PostgreSQL and vector stores for embedding-based lookups.
• Apache Airflow DAG pipelines for feature engineering, training refresh, and scheduled re-scoring; MLflow for experiment tracking and model lineage.
• Guardrails-first design — every production model ships with a monitoring harness (drift, latency, approval-rate deltas), fallback rules for low-confidence cases, and explainability artifacts appropriate for regulated review.
• Privacy-by-design — personally identifiable fields are pseudo-anonymized at ingest; sensitive model paths run with on-prem inference to satisfy data-residency requirements.

Impact

• 500K+ monthly transactions processed at 92% accuracy
• 25% uplift in targeted loan-product uptake
• 40% faster underwriting review cycles
• 30% fewer fraud false-positives
• Systems delivered as part of engagements contributing to multi-year banking client relationships.