MS in Data Science @ NYU (GPA 3.9/4.0) · Published Researcher — 3 peer-reviewed papers
Machine learning for credit risk, reinforcement learning, and LLM systems
I build ML systems with measurable results, calibrated probabilities, backtested pipelines, benchmarked agents.
XGBoost default-probability model on 1M+ firm-year records with isotonic calibration — decision-usable, auditable outputs for corporate default prediction.
- Walk-forward AUC 0.837 on time-validated data, using finance-grounded features (Altman Z, leverage, liquidity)
- Packaged as a CLI scoring harness with persisted artifacts (model, calibrator, metadata) for reproducible inference
Can an LLM replace a human reward engineer? An automated Claude loop that designs and revises PPO reward functions on Super Mario Bros.
- Matched 74.5% of an expert's hand-tuned 8-iteration result in just 5 fully automated rounds
- Revises rewards using training diagnostics (entropy, episode length, explained variance) as feedback
- Caught reward-hacking exploits and two evaluation bugs via methodology audits; rebuilt the comparison on a leak-free criterion
Gender bias analysis across 90k+ instructor reviews with rigorous statistics.
- Surfaced a statistically significant pro-male rating bias (Mann-Whitney U on Bayesian-adjusted ratings, bootstrap CIs) — and showed it is practically negligible (Cliff's Δ ≈ 0.04)
- Found 18 of 20 student tags significantly gendered; modeled ratings (test R² = 0.79) and “pepper” status (AUROC = 0.94)
Air-Pollution-Prediction-with-Advanced-Preprocessing-and-Deep-Ensemble-Learning — 📄 Atmospheric Pollution Research, 2025
BiLSTM + XGBoost weighted ensemble forecasting hourly PM2.5 from a 48-hour lookback.
- R² = 0.89 on 30k+ samples, outperforming individual-model baselines under strict temporal validation
- Reusable preprocessing pipeline with PPCA imputation and robust scaling
Solving-the-0-1-Knapsack-Problem-Using-the-LAB-Algorithm — 📄 Springer Handbook of Formal Optimization, 2024
LAB (Leader–Advocate–Believer) metaheuristic with constraint-repair and stagnation-triggered perturbation.
- Reached known optima on 19/20 single-knapsack and 20/30 multidimensional WEISH benchmarks (remaining gaps < 0.7%)
- Shipped as a reproducible Python package with automated tests and CLI benchmarking
Scribee ♿
Sign language & Braille translation app for accessibility — led the team shipping it end to end.
- Real-time hand-sign recognition (MediaPipe + TensorFlow) integrated with speech recognition and a Streamlit UI
- Narkhede, G.G., Poonawala, M., et al. Air Pollution Prediction with Advanced Preprocessing and Deep Ensemble Learning. Atmospheric Pollution Research, 2025.
- Poonawala, M., Kulkarni, A. Solving the 0-1 Knapsack Problem Using LAB Algorithm. Handbook of Formal Optimization, Springer, 2024.
- Poonawala, M., et al. LoRa-Based Farm Monitoring System. ICT Analysis and Applications, Springer, 2023.
Statistics & Modeling: Hypothesis testing · Bayesian inference · bootstrap resampling · walk-forward validation · probability calibration · time series