I build ML systems that deploy to production, not just live in notebooks.
Currently building VEKTOR — a from-scratch semantic search engine with a custom HNSW index, hybrid dense+sparse retrieval, and a self-improving cross-encoder reranker. Also working on healthcare AI for early disease detection from clinical time-series data.
- End-to-end ML — from raw data ingestion to production REST API
- Accuracy without scalability is just a prototype — I build for both
- B.Tech CSE (AI/ML) @ KIIT · IEEE-published researcher · 3 industry internships
| Project | What it does | Result |
|---|---|---|
| VEKTOR | From-scratch semantic search engine — pure-NumPy HNSW index, hybrid dense+sparse retrieval via RRF, self-improving cross-encoder reranker, async FastAPI + Redis + Prometheus + Docker | Sub-50ms p99 latency, 1M+ docs, recall@10 ≥ 0.95, +8% MRR from reranker |
| NEO SEPSIS | Early neonatal sepsis detection from 40K+ NICU time-series (MIMIC-III) — XGBoost + LSTM ensemble, SMOTE for class imbalance, SHAP/LIME per-patient explanations | AUROC 0.921 — comparable to specialist-physician performance |
| DUALFUSION | Thyroid ultrasound segmentation — hybrid CNN–Transformer dual-encoder (Swin-Transformer global attention + ResNet local features + UNet++ decoding), benchmarked on DDTI, TN3K, TGT3 | Dice 0.90, 92%+ pixel accuracy; outperforms ViT baselines by 7–12% · IEEE 2025 |
| SOLARVERSE | Real-time solar flare & CME forecasting — 25+ features engineered from live NOAA GOES-16 satellite data (X-ray flux, magnetic field vectors, sunspot indices), multi-horizon XGBoost models | 6h/12h/24h forecasts with automated alerts; Flask REST API on GCP Cloud Run |
| Song Popularity | Spotify track popularity classifier on 114K+ tracks — XGBoost on audio features (danceability, tempo, valence, energy) with SMOTE resampling | FastAPI /predict endpoint + HTML/JS frontend; F1, AUC-ROC evaluation |
| ClimateVision | Open-source ML platform for satellite-based environmental monitoring — deforestation, ice-melt, and flood detection from Sentinel-2 / Landsat imagery | Python + TypeScript; active open-source contribution |
| Map My Way | SIH trip tracker — GPS trajectory analysis, business/leisure classification via logistic regression, OCR expense tracking, Phi-3 Mini chatbot, Go/Echo REST API | Multi-modal AI integration |
| Jurify | AI legal document simplifier — BART summarization, sentence-transformers Q&A chatbot, multi-format upload (PDF, DOCX, TXT), side-by-side comparison | Production-ready NLP pipeline |
Computer Vision · XGBoost · Scikit-Learn · OpenCV · Flask · FastAPI · REST APIs
Deep Learning · NLP · Transformers · GCP Cloud Run · Pandas · NumPy · Docker · Redis