Résumé

On-site

• Engineered a Flask + Oracle Database system spanning 16 modules and 42 REST routes to digitise the full industrial-asset lifecycle (inventory, grinding, inspection, assembly) across a 17-state roll-status workflow tracked over 12 relational tables.
• Built an Excel ingestion pipeline enforcing 40-column schema validation and 26 required-field checks, with transactional rollback and full audit-trail backups on every submission, plus business-rule validation (6-roll-per-stand capacity limits, duplicate-type position checks) for stand assignment.

Remote · two virtual internships

• Built and deployed an NLP-based college FAQ chatbot (Flask, scikit-learn, NLTK) covering 75 intents, using WordNet-based augmentation to expand training data to 3,200+ examples; benchmarked 5 classifier families with GridSearchCV, achieving 92.7% test accuracy with a tuned Random Forest.
• Built a hospital readmission risk prediction system (Streamlit, XGBoost, SHAP) trained on 100K+ patient records and 13 clinical features, achieving 85.2% accuracy and 88.9% AUC-ROC, with SHAP-based explainability and an interactive analytics dashboard.

On-site

• Studied Mumbai's EMU railway operations, maintenance workflows, hardware infrastructure and safety protocols alongside engineering teams, building a working understanding of large-scale industrial systems.

• Built a full-stack heritage-exploration platform (React 19, Express 5, MongoDB, Three.js, Socket.IO) integrating 9 external REST APIs across 126 heritage sites (36 UNESCO) with 780 quiz questions, plus a Groq-hosted LLM assistant for heritage Q&A, safety guidance and structured AI trip planning.
• Engineered a fully client-side MediaPipe gesture-recognition pipeline (9 gesture classes at 20 FPS), validated against the original implementation across 20,000 test poses with zero mismatches, replacing a legacy server relay that cost ~3.2 GB/hour per user.
• Designed a two-tier (in-memory + MongoDB) caching architecture with per-resource TTLs and API-key pooling, cutting a live lookup's response time from 8.7 s to 2 ms, and built a real-time 1v1 quiz-duel mode (4 modes, 10 rounds) over Socket.IO.

• Built a Unity 6 (C#) VR/desktop chemistry-lab simulator (~110 scripts, ~40,000 LOC) featuring 8 simulated inorganic reactions with a custom free-hand quantity-validation engine that detects overdose, underdose, wrong-order and heating-duration errors in real time.
• Developed an in-lab AI chemistry assistant (Convai API, voice + text) grounded in live experiment measurements, with an offline rule-based fallback for uninterrupted use, plus live in-engine 3D molecular animations (bond breaking/forming, electron transfer) for all 8 reactions.
• Built a separate assessment mode with 18 timed tasks (8 practical, 10 theoretical) and a coin-economy ranking system; validated the full simulation with 226/226 passing automated tests.

• Engineered a full-stack hospital operations platform (Django REST Framework, React, MySQL) with 11 feature modules across a 33-table schema, serving 4 role-based dashboards (Doctor, Nurse, Admin, Receptionist) for real-time bed occupancy, OPD queue management and inter-hospital capacity sharing.
• Trained and deployed 6 scikit-learn models across 3 prediction domains (OPD wait-time regression, inventory stockout classification + demand-depletion regression, and billing profit/loss prediction), plus a rule-based engine linking live OpenWeatherMap/AQI data to medicine-demand forecasting.
• Integrated a secure Razorpay payment gateway (order creation, HMAC-SHA256 webhook signature verification) for end-to-end billing and payment confirmation.