msc math @ iit delhi · building ML systems that ship — from data to deployment
- Hybrid Architectures: Bridging classical tree-based models (XGBoost/CatBoost) with deep metric learning (Siamese CNNs) and Transformer-based NLP (DistilBERT).
- Robust Evaluation: Designing leak-free splitting protocols (skeleton-aware) and auditing open datasets to identify data leaks.
- Production Deployment: Wrapping models in interactive Streamlit and Hugging Face spaces for live inference.
-
✍️ Signature Forgery Verification
- Motive: Build a writer-independent signature verifier that generalizes to unseen writers, while identifying and mitigating data leakage in standard ICDAR datasets.
- Tech: Siamese CNN & fine-tuned EfficientNet-B0 + online batch-hard mining.
- Result: Achieved 0.986 ROC-AUC and mitigated a systemic dataset leak to establish a robust writer-independent evaluation protocol.
- HF Space Demo
-
- Motive: Audit and detect CCPA-illegal manipulative UI patterns (dark patterns) on e-commerce platforms to protect consumers.
- Tech: Fine-tuned DistilBERT vs. Optuna-tuned classical models (XGBoost/SVC).
- Result: Achieved 0.797 macro-F1 under strict skeleton-aware, leak-free splits for 13 compliance categories.
- Classical Demo · Transformer Demo
-
- Motive: Quantify skill premiums and cost-of-living adjustments to help Indian tech workers negotiate salary transparency.
- Tech: CatBoost regressor optimized with Optuna.
- Result: Achieved 0.8753 R² on 88K tech worker profiles, mapping compensation across roles and cities.
- Streamlit Demo
-
- Motive: Predict flight delays using strictly pre-departure features to serve as a traveler warning system.
- Tech: Benchmark of 6 estimators; deployed pre-departure XGBoost model.
- Result: Achieved 0.773 ROC-AUC for binary flight delay classification.
- Streamlit Demo
-
- A zero-dependency, rapid-lookup reference tool and decision wizard to help developers select optimal models and validation methods.
- Live Site
-
- It provides a quick-reference concept companion and pipeline integrity diagnostic quiz for NLP practitioners.
- Live Site
┌─────────────────┬────────────────────────────────────────────────┐
│ Classical ML │ scikit-learn · Optuna · XGBoost · CatBoost │
├─────────────────┼────────────────────────────────────────────────┤
│ Deep Learning │ PyTorch · TensorFlow · Keras · Hugging Face │
├─────────────────┼────────────────────────────────────────────────┤
│ NLP / Text │ spaCy · NLTK · DistilBERT │
├─────────────────┼────────────────────────────────────────────────┤
│ Deploy & MLOps │ Streamlit · Hugging Face Spaces · GitHub Pages │
└─────────────────┴────────────────────────────────────────────────┘