M.Tech Artificial Intelligence student at IIT Roorkee. I build ML systems end-to-end — retrieval, agentic pipelines, and production deployment and I'm currently interviewing for Data Science / MLE / Applied Scientist / GenAI-LLM Engineer roles.
🔭 Core interest: multimodal ML (vision + text), retrieval-augmented systems, and taking research ideas to production-shaped code.
Enterprise Agentic Multimodal RAG Platform
A from-scratch RAG platform ingesting PDFs, PPTs, Excel, images, tables, and diagrams via multimodal embeddings (CLIP, MiniLM). Hybrid retrieval (BM25 + Milvus dense search, reciprocal-rank fusion), a Neo4j knowledge graph with NER-based entity extraction, and a Groq-backed (Llama 3.3) agentic layer.
Python · FastAPI · Milvus · Neo4j · PostgreSQL · Redis · Docker · Kubernetes · Sentence-Transformers · spaCy · RAGAS · DeepEval
CacheRAG
A KV-cache optimization framework for enterprise RAG that weighs the business cost of evicting critical evidence, not just cache-hit rate. Five scoring signals feed a two-level fractional knapsack budget allocator, evaluated with a custom CERR metric — outperforms all baselines at 35–50% compression in simulation. Ships with a FastAPI service, interactive demo, and Docker setup.
HuggingFace Transformers · vLLM · FastAPI · Docker
Multimodal Product Search & Recommendation Engine
A production-shaped recommendation system over the Amazon Reviews/Products dataset — SQL ingestion, EDA, TF-IDF + image-feature embeddings, FAISS retrieval with re-ranking, served via FastAPI with a demo UI, containerized with Docker.
Python · FAISS · FastAPI · Docker · scikit-learn
Multimodal RAG for Scientific Literature (M.Tech Thesis) Joint text-image retrieval building on ColPali-style late-interaction (MaxSim) scoring with a PaliGemma backbone, benchmarked against ViDoRe, with RAGAS-based faithfulness evaluation.
Python PyTorch FastAPI FAISS Milvus Neo4j Docker Kubernetes Sentence-Transformers spaCy PostgreSQL Redis