Third-year Computer Engineering student who builds RAG pipelines, deep-learning models and RL systems from scratch — hybrid retrieval, reranking, evaluation and all the math underneath.
The human behind the commits — what I build, what I'm learning, and where the journey is headed.
I'm a third-year Computer Engineering student at Ganpat University (UVPCE), specialized in AI/ML. I don't stop at calling APIs — I build the machinery: hybrid retrieval with BM25 + dense vectors fused via RRF, cross-encoder reranking, RAGAS evaluation loops, CNNs with Grad-CAM interpretability, and value iteration written from the Bellman equation up.
I've contributed to HandsOnLLM — the companion repo to O'Reilly's Hands-On Large Language Models — and completed the Stanford / DeepLearning.AI ML Specialization, implementing every algorithm from scratch in Python.
Currently hunting for a Summer 2026 internship where I can point all of this at real-world problems.
Ganpat University (UVPCE), Mehsana — DSA (Java), OS, DBMS, Networks, Linear Algebra, Probability & Statistics.
Stanford / DeepLearning.AI ML Specialization (credential BDRUK7BXSI0S) — every algorithm from scratch. Shipped MEd X and the Fake News Detection System.
Built the Multi-Document RAG Pipeline (Ask-My-Paper). Pull request to HandsOnLLM, O'Reilly's companion repo. Hunting a Summer internship.
Graduation loading — sem 6 of 8.
Languages, frameworks, retrieval systems — and everything that keeps them honest.
Retrieval, vision, and language — three systems where I owned the whole pipeline, from ingestion to evaluation.
End-to-end RAG with citation tracking across documents. Hybrid retrieval — ChromaDB/FAISS dense + BM25 sparse, fused via Reciprocal Rank Fusion — then FlashRank cross-encoder reranking before generation. PyMuPDF ingestion with adaptive chunking; evaluated on RAGAS.
Deep-learning classifier for chest X-rays using DenseNet121 transfer learning on the NIH dataset. Data augmentation to fight class imbalance, plus Grad-CAM activation maps so a clinician can see exactly where the model is looking — interpretability first.
NLP classifier that separates real headlines from fake ones using a BiLSTM + BERT hybrid architecture. Text preprocessing and embedding pipeline feeding a bidirectional LSTM on top of BERT representations, served via a FastAPI inference endpoint.
Stats, the contribution graph, and my latest repos — pulled straight from the GitHub API on every visit.
Recruiting for a Summer 2026 AI/ML internship, or just want to argue about chunking strategies? My inbox is open — I usually reply within a day.