Machine Learning Engineer with 3 years building credit-risk decisioning systems end-to-end (modeling → pipelines → AWS production) at a listed SME-lending NBFC and its consumer-lending subsidiary.

• 🎓 MSc in Data Science @ Nanyang Technological University, Singapore - College of Computing and Data Science (Aug 2026 - Jun 2027)
• 💼 Recent work: default-prediction AUC 0.65 → 0.75 via a counterparty GNN over 150M+ transactions, an early-warning system at AUC 0.87, and a 74% model-serving cost cut
• 🔬 Into GNNs, time-series, model interpretability, efficient ML, and audio - with a soft spot for Kaggle leaderboards 🎹

📌 Seeking Singapore-based internships alongside the MSc: 16 h/week during term (classes run evenings) and full-time during vacations, under MOM's student work-pass exemption - no separate work pass required.

MyShubhLife (now GROx Technologies), UGRO Capital Group - consumer-lending fintech · Bengaluru, India

• Shipped an early-warning system predicting borrower default 30 days ahead: two-segment ensemble (clean vs. failed-auto-debit borrowers) on recency features, test AUC 0.87; risk tiers drive collections prioritisation, cutting roll-forward rates 8% over one quarter.
• Raised default-prediction test AUC 0.65 → 0.75 over a logistic baseline with a counterparty graph neural network (shared counterparties propagate risk), trained on 37K borrowers / 100K bank statements / 150M+ transactions and served within production latency budgets.
• Owned the Gro Score credit-scoring API on AWS behind live lending decisions - automated validation suites, feature-ablation analyses, and schema and completeness checks on third-party bank-statement payloads that caught truncated or malformed JSON before it reached the model, triggering upstream re-pulls instead of silent mis-scoring.
• Built 7-day cash-flow forecasters on 180-day transaction histories (MAPE 12.5%, stable backtests across cohorts); productionised the feature and evaluation suite used across credit-limit experiments and quarterly underwriting policy reviews.

UGRO Capital - listed SME-lending NBFC · Mumbai, India

• Containerised TensorFlow scoring models and migrated serving from a single 8-core EC2 instance to AWS Lambda: infra cost down 74%, end-to-end scoring 60s → 28s, and removed the capacity ceiling behind peak-hour failures. All UGRO lending products routed through this API, underwriting applications converting to ~US$30M (INR 250 Cr) in monthly disbursals.
• Rebuilt the tax-filing (GST) default-prediction model after portfolio review found the incumbent at test Gini 16 - logistic regression, random forest, and ANN with spline smoothing over proprietary filing-pattern features - lifting test Gini 16 → 36 (train 38; minimal generalisation gap) on 30+ DPD at 12 months; features reused across multiple lending products.
• Built the pre-release monitoring and regression-test suite (schema checks, feature-sanity tests, latency SLO gates, drift alerts, automated reports) that caught regressions before release and gated the team's frequent production deploys.

Languages & Core

Python (NumPy, pandas, Polars) · SQL · C++

ML / Deep Learning

Time-series forecasting · backtesting · credit-scorecard modelling · feature engineering · model validation

MLOps & Infrastructure

AWS (Lambda, EC2, S3, CloudWatch) · Docker · PostgreSQL · Git · FastAPI · CI/regression · model & drift monitoring