Fraud detection, risk analytics, and identity & payments risk, built on SQL/BigQuery and grounded in hands-on investigations.
5+ years across a U.S. digital bank (Monzo US), a global crypto exchange (ProBit Global), and banking financial-crime roles (HSBC, SMBC, Bank of Communications). M.S. in Business Analytics, USC Marshall.
- Fraud detection & rule quality: owning rule portfolios, tuning signal combinations and thresholds against investigation outcomes, and cutting non-actionable alerts while known fraud is still caught
- Identity & onboarding risk: evaluating vendor scores and onboarding decisions against downstream fraud, balancing fraud capture, review volume, and customer friction
- Payments risk & root-cause analysis: tracing anomalies to their source in BigQuery and partnering with Engineering through remediation
- Metric integrity: rebuilding loss methodology, tracing risk metrics to source tables, and validating the numbers behind board and leadership reporting
- Investigations → controls: account-takeover and linked-account analysis with device, IP, behavioral, and transaction signals, turned into scalable onboarding controls
Professional work. No confidential data is published here.
Monzo US
- Rebuilt U.S. fraud-loss methodology in BigQuery across three source datasets, aligning definitions with Monzo UK; fixed loss classification errors (including won chargebacks counted as losses) and revised reported fraud losses down by 35%
- Backtested a vendor fraud score across 27K+ onboarding applications with Data Science; recommended targeted verification and manual review rather than automatic rejection
- Traced a negative-balance spike to an ACH return-file issue blocking ~250 valid returns; partnered with Engineering on remediation
- Validated seven fraud and operations metrics across BigQuery and Looker, finding that fraud cost per active customer needed a 26% upward correction; supported an overnight board risk update by recalculating 12 months of fraud-loss bps (exact match)
ProBit Global
- Owned and refined 20+ active fraud rules; reduced non-actionable alert volume by 35% within a targeted rule set while verifying that known fraud was still detected
- Expanded account-takeover investigations to linked accounts using SQL; evidence-based restrictions reduced at-risk withdrawal exposure by 43% across investigated accounts
- Translated linked-account patterns into device, IP, behavioral, and session-based onboarding controls
| Project | What it shows | Result |
|---|---|---|
| Application Fraud Detection | Placeholder cleanup to prevent false identity links, 1,500+ linkage/velocity features, KS + forward selection, 5-model comparison | 61.2% of fraud captured in top 3% (OOT) |
| Card Transaction Fraud Detection | 2,600+ behavioral and velocity features across 0–60 day windows, LightGBM | 57.2% of fraud captured in top 3% (OOT) |
| NYC Property Anomaly Detection | Unsupervised scoring with no labels: peer-normalized features, PCA distance + autoencoder, explainable review queue | 1.04M records ranked; top-10K queue |
All three are academic projects on course or public data, not production systems.
Analysis: SQL (BigQuery, CTEs, window functions) · Python (pandas, scikit-learn, LightGBM) · dbt (code review) · Looker · Tableau AI-assisted workflow: Claude Code · Cursor · Gemini CLI for query work, with independent recalculation to validate report figures Investigation: Chainalysis · World-Check · Jira