Data Science · Machine Learning · Open to Roles
"Models that ship - with explainability and a business case."
komal = {
"role": "Data Scientist · ML Engineer",
"based_in": "Delhi, India",
"education": "MCA '26 · Open to roles",
"focus": ["applied ML", "federated learning", "cost-sensitive modeling"],
"ml": ["XGBoost", "scikit-learn", "TensorFlow", "SHAP", "SMOTE"],
"ships": ["Streamlit", "Flask", "Git"],
"currently": "building a 200+ row Indian DS/ML job-market dataset"
}
| Project | What it does | Stack | Live |
|---|---|---|---|
| 🌸 FedFlower | Federated-learning CNN that beats the centralized baseline by +3.90 pts (91.82% Top-1) at ~1/5 the per-client storage. Privacy-preserving; deployed on Android, on-device. | TensorFlow Federated Learning Android |
— |
| 🧬 smote-reproduction | From-scratch NumPy reproduction of SMOTE (Chawla et al., 2002) - rebuilds the paper's ROC/AUC protocol on Pima Indian Diabetes (1.87:1) and Mammography (42:1). AUCs land within ~0.01–0.03 of the paper across both; from-scratch SMOTE matches imbalanced-learn to 0.001. | NumPy scikit-learn Paper Reproduction |
— |
| 💳 Credit-Decision-Intelligence-System | Loan approval optimized for profit, not accuracy — per-loan cost-matrix threshold (not 0.5). ~$49M more holdout profit than the conservative baseline. | XGBoost SHAP Streamlit |
▶ Demo |
| 🏥 health_insurance_claim_prediction | Predicts claim costs at R² 0.960 / MAE $1,053; SHAP surfaces cost drivers (smoking, heart disease, age/BMI) for underwriters. | XGBoost SHAP Streamlit |
▶ Demo |
"Data is the new oil - but refining it is the hardest part."