ML Engineering · Open-Source Scientific Software · Applied Research
B.Tech Mineral & Metallurgical Engineering · IIT (ISM) Dhanbad · 2024–2028
I'm an undergraduate student in Mineral & Metallurgical Engineering at IIT (ISM) Dhanbad, self-directing a focused transition into applied ML and open-source scientific software infrastructure. My background gives me domain ownership — I understand the physical systems my models are predicting, not just the loss functions.
My work lives at the intersection of rigorous ML engineering and real industrial problems: anomaly detection in time-series data, multi-task computer vision for industrial inspection, probabilistic simulation, and LLM-integrated process control systems.
| Area | Work |
|---|---|
| Anomaly Detection | HampelFilter for sktime (PR #9757) · Blast furnace PCA anomaly pipeline |
| Computer Vision | ScrapScan — multi-task EfficientNet-B0 · ONNX in-browser · Score-CAM explainability |
| Probabilistic Modeling | FIFA WC 2026 Monte Carlo engine — 50K iterations, Dixon-Coles Poisson |
| LLM Integration | Qwen2.5-7B process control layer translating SHAP outputs → operator guidance |
| ML Infrastructure | neural-lam config validation (PR #529) · scikit-learn-compliant estimator design |
| Evaluation Research | MetallurgyBench — LLM benchmark for industrial materials science reasoning |
sktime — Python's most widely used time-series ML library
- PR #9757:
HampelFilteranomaly detector — check_estimator compliant, MAD-based thresholds, configurable window size, full validation suite pass
neural-lam — PyTorch neural weather prediction framework
- PR #529: Configuration schema validation layer — automated checks preventing invalid parameters from reaching GPU allocation, eliminating wasted compute
| Project | What it does | Key result |
|---|---|---|
| Blast Furnace Slag Viscosity Controller | XGBoost+CatBoost+DNN ensemble + Qwen2.5-7B LLM, deployed on HF Spaces | ~0.97 Test R² |
| FIFA WC 2026 Prediction Engine | 50K Monte Carlo iterations, Dixon-Coles, vectorized 50min→60sec | 56.5% accuracy on 69 live matches |
| Credit Card Fraud Detection | LightGBM+XGBoost, SHAP, F2-optimized threshold | ROC-AUC 0.9845 |
| F1 Pit Stop Duration Prediction | 12-model stacked ensemble (XGBoost/LightGBM), time-aware CV, SHAP interpretability | Identified 2009-10 covariate shift · sklearn GB outperformed XGBoost |
| MetallurgyBench | LLM evaluation benchmark for industrial materials science reasoning | Design + task construction phase |
Languages: Python · C · C++ · SQL
ML & DL: PyTorch · TensorFlow · XGBoost · LightGBM · scikit-learn · Optuna · SHAP · ONNX
Computer Vision: EfficientNet · OpenCV · Score-CAM
LLMs: Hugging Face Transformers · Qwen2.5 (inference + explanation pipelines)
Data: NumPy · Pandas · SciPy · Matplotlib · Seaborn
Tools: Git · Gradio · Hugging Face Spaces · Jupyter · Google Colab
- 🏢 ML Engineering Intern @ FlyRank AI (Jul 2026 – Present)
- 🔬 AI/ML Researcher @ Dubai Computer Science Society (Jul 2026 - Present)
- 🌏 Data Science Fellow @ Matsuo-Iwasawa Lab, University of Tokyo (Apr 2026 – Present)
- 📖 Building MetallurgyBench — the first LLM evaluation benchmark for industrial materials science process reasoning
- 🥇 Top 25 — IIM MATRIXx 2026 Hackathon (IIT Hyderabad) · ScrapScan · JSW Steel R&D collaboration offer
- 💻 CodeChef max rating 1526 · Global rank 646 in Starters 227 (40K+ participants)
- 🧠 AlgoUtsav 2026 — Rank 151 / 2,500+ teams (NIT Rourkela)
- 🏐 Gold Medal — General Championship Inter-Hostel Volleyball Tournament, IIT (ISM) Dhanbad (libero)
To drive cutting-edge AI advancements by securing an MS in Machine Learning, publishing impactful research, and contributing to large-scale open-source ecosystems.