At Tata Consultancy Services, I built a nine-step Node.js LLM pipeline for a Nissan TARA tool, turning architecture documents into threat-analysis and risk reports in 30–60 minutes on local hardware.
On an eight-person team, I translated the team-defined TARA workflow into pipeline steps and validation checks. Across 34 runs, the tool generated 1,400+ threat scenarios at $0 API cost.
I built the retrieval layer with the RAG team over ISO/SAE 21434, UN R155, and VDA materials, alongside an expert-validated knowledge base and knowledge graph. The system retrieved the right passage in the top three for 84% of hand-labelled queries.
For a second OEM, I built a multimodal research agent that searches web, PDF, and image sources, cites claims, and produces slide decks. I also fine-tuned Mistral 7B with QLoRA and served GGUF through Ollama; in my LedgerLens project, I built a FastAPI receipt and invoice extraction service with validation, review workflows, and 181 tests.

