Pranjal Chaubey

I build AI systems that go into production and stay there.

I work on Tata 1mg's AI Transformation team, where the job is to walk into a department, map what people actually do, and rebuild that process around AI. That framing matters more than any framework, because the hard part is rarely the model. It is the messy operational data underneath, the entity that appears under six different names, the process nobody has revisited in a decade.

Some of what I have built: Optimus, an enterprise copilot over a centralised RAG store, grounded in ReAct with RAPTOR and CRAG retrieval on Qdrant, FastAPI, Redis and OpenAI. We shipped it inside the chat platform employees already used rather than as one more tab. It serves 6,000+ employees and absorbs roughly 30% of the query load across 20 HR business partners.

Webassure, product intelligence over a drug catalog. LLM structured extraction producing 64 fields from low-quality product images, with hybrid semantic search resolving each extracted entity against existing database IDs across 800,000+ SKUs instead of creating duplicates.

Compliance moved from around 70% to over 98%, and onboarding scaled to 10,000+ SKUs a month. A real-time multimodal audit pipeline for insurance pre-medical calls: ElevenLabs transcription plus Gemini discrepancy detection across 186 questions, flagging only the exceptions to a human reviewer and auto-submitting the rest. It went from 300 to 1,500+ calls a day at roughly 98% accuracy, with turnaround time down 67%, PII-compliant end to end. Multi-signal retail search fusing BM25, dense embeddings, phonetic and fuzzy matching through Reciprocal Rank Fusion with LLM reranking, across 15 stores. I also own enterprise AI governance where I work, using Langfuse for LLM observability, MLflow for tracking and Skyflow for PII handling, and I run the technical evaluation of AI vendors before we adopt them.

Outside work I keep building. MAPLE is an AI-native Android finance app running an agentic loop of memory, reflection, user model, judgment and action, with an MCP-based tool layer. I have also trained a multi-stage neural collaborative filtering recommender, built an MCP search server, and a text-to-SQL system over multi-source data. Most of it is public at github.com/ddcrpf.

I studied Artificial Intelligence and Machine Learning at Nitte Meenakshi Institute of Technology (B.E., 9.07/10). What I want next is a small team where I can own systems end to end, close to the problem, with the autonomy to fix what breaks. Gurugram, India. prnjlchaubey@gmail.com