Rohan Dhupar

Built a voice AI platform (WebRTC + GPT-4o) for knowledge transfer, creating a searchable RAG pipeline (ChromaDB) to replace manual offboarding. Designed microservices backend (FastAPI + React + Node.js) with Docker. Stack: Python, Azure OpenAI, GPT-4o, RAG, ChromaDB, WebRTC, Docker.

Automated annotation pipeline with Cog-VLM and LLMs, cutting manual data creation by 66%. Built knowledge-graph-enhanced time-series models improving product-usage prediction accuracy by 28%. Mentored 2 ML engineers and introduced experiment tracking and staged model review processes.

Delivered real-time document digitisation pipeline (DBNET + SAR + LLaMA2) achieving ~95% accuracy, reducing retrieval from minutes to under 15 seconds. Replaced BM25 with Milvus vector retrieval, improving precision and increasing document throughput by ~33%.

Built real-time SIM fraud detection combining GPT-2 text signals with IR-SE facial verification and YOLOv5 document checks, catching 25% of fraudulent registrations. Replaced BM25 with Milvus ANN search, reducing false-negative rate by 22% with no added latency.

Fine-tuned BERT and T5 on legal corpora, improving termination/renewal clause extraction by 42% and party identification by 48% over rule-based baseline. Implemented batched GPU inference, reducing contract review time from ~40 min to under 4 min per document.