Senior Data Scientist with 4+ years turning production ML into measurable outcomes: $1.7M+ in business value delivered across fraud detection, NLP, forecasting, and risk analytics, most recently at AXA Group Operations (Paris & Kuala Lumpur). I own models end-to-end: from a messy CSV to a monitored, containerized scoring API.
| 🎯 Focus | Fraud & Anomaly Detection · NLP · Forecasting · MLOps |
| 🧰 Toolkit | Python · PyTorch · TensorFlow · scikit-learn · FastAPI · Docker |
| 📍 Based in | Bangalore, India (open to relocation & remote) |
| 🗣️ Languages | English · Hindi · Bengali |
| 🌐 Portfolio | hridaysaha.com |
| Impact | Where |
|---|---|
| ⚡ 94.2% F1 real-time fraud model on a 1M+ txns/day platform at sub-100ms; $1.5M+/yr fraud loss prevented | AXA Group Operations |
| 🔍 93% Top-1 / 99% Top-3 NLP audit-mapping system; 96% less manual effort, ~$180K/yr saved | AXA Group Operations |
| 📊 Delivered a $1.2M analytics portfolio at 98% on-time, leading a team of 3–5 across 40+ stakeholders | AXA Group Operations |
| 🎓 Advised 60+ applied ML projects; trained 150+ professionals (92% completion) | The Sparks Foundation |
Languages & Databases
Machine Learning · Deep Learning · NLP
LLM & RAG Frameworks
Vector Databases
Cloud & App Frameworks
Deployment & MLOps
Visualization & BI
Projects spanning the full ML spectrum (fraud · causal inference · GenAI · NLP · forecasting · MLOps), each built with production rigor and honest evaluation.
| Project | What it does | Stack |
|---|---|---|
| 🔗 Fraud Detection Platform | Point-in-time fraud scoring on the IEEE-CIS transaction stream: feature store with bit-identical online rows, cost-asymmetric thresholds, PSI drift monitoring and a gated retrain loop. Measures the leakage each shortcut buys (a random split alone inflates PR-AUC by +0.26); 37.6 ms median scoring path. | XGBoost · Redis · FastAPI |
| 🔗 DriftWatch | Model-agnostic ML monitoring service that detects feature/prediction drift with delayed-label support and effect-size-based alerting. 219 passing tests, fully containerized, seeded fault scenarios verified in CI against a real Postgres. | FastAPI · PostgreSQL · Docker |
| 🔗 Uplift Targeting | Per-customer treatment-effect estimation turned into a budgeted, dollar-valued targeting policy. Meta-learners vs. uplift forests on 14M Criteo rows, where uplift clears the response baseline by 4–6% Qini (p < 0.001) and honestly doesn't on 64k Hillstrom customers. 130+ tests, live Streamlit demo. | scikit-uplift · LightGBM · Streamlit |
| 🔗 Insurance Policy RAG | Grounded Q&A over insurance policies: hybrid retrieval with clause-level citations, 34/34 citations verified and 3/3 correct refusals when the answer isn't in the document. Runs fully local via Ollama. | SBERT · FAISS · BM25 · Ollama |
| 🔗 CVE→CWE Semantic Mapping | Auto-classifies vulnerability text to weakness categories via staged retrieve-then-rank: 87.4% top-1 / 96.1% top-3, with SHAP explanations on every prediction. | SBERT · LightGBM · SHAP |
| 🔗 FinSight | Forecasts next-quarter revenue, operating income & free cash flow from SEC + FRED data across 117 companies: leakage-safe rolling-origin validation, per-company MASE, and the discipline to ship the seasonal-naive baseline for FCF when no learned model beat it. | LSTM · LightGBM · Prophet |
- 🎓 M.Sc. Data Science & Business Analytics, Asia Pacific University (APU), Malaysia
- 🎓 B.Tech. Computer Science & Engineering, MCKV Institute of Engineering, India
- 📜 Certifications: IBM Data Science Professional · SAS Data Science Professional · Microsoft Power BI Analytics
- 🏆 AXA Innovation Excellence Award (2023) · AWS Build Malaysia Semifinalist (top 10% of 500+)
I'm open to Senior Data Scientist / ML Engineer roles: full-time, remote, or relocation.
