Divyaraj Singh Chundawat

Architected the core LangGraph multi-agent backbone of Nexus SAM from scratch: a 13-node StateGraph with TypedDict shared state, custom reducers, conditional routing, and clarification and follow-up feedback loops driving each RAG turn from query classification through multi-source retrieval to response synthesis. Built human-in-the-loop conversation flows with LangGraph interrupt() / Command(resume), persisted across HTTP request boundaries via a MongoDB checkpointer keyed on conversation id, so the agent pauses for clarifying questions and resumes exactly where it left off. Owned the Savitar RAG pipeline end to end: Elasticsearch hybrid search (BM25 + semantic), per-query sub-query breakdown, multi-persona YAML prompting, Pydantic structured-output parsing, and inline citation anchoring threaded from retrieval to final answer; added a higher-recall Deep Search mode. Built a parallel source-fan-out executor (asyncio.gather) that dispatches across five maritime knowledge domains through a typed tool-node dispatch table and merges per-source results into unified structured responses. Shipped an email-intelligence layer over Microsoft Exchange: LLM reply suggestions, a 400+ line action-tagging classifier, awaiting-reply detection, and a Microsoft Graph API ingestion pipeline classifying 20,000+ port / vessel records. Built event-driven automation workflows for machinery-failure and incident tracking: a hierarchical LLM classifier detects relevant emails, extracts structured data, raises real-time alerts, and maintains per-vessel lifecycle tracking in MongoDB, all on a modular, pluggable FastAPI service that accepts new workflows without re-engineering the core. Raised the domain RAG pipeline from 60% to 96% accuracy (chunking redesign, reranking, injected domain knowledge) and cut end-to-end chatbot latency 72% (180s to 50s) via cost-performance LLM selection; benchmarked and served models on NVIDIA H200 with vLLM. Enforced agent reliability with Pydantic structured output, hallucination-reduction fixes, and an error-envelope pattern; instrumented tracing, latency profiling, and output evaluation with LangSmith; tuned YAML prompts for 7+ agents across 5 tenant personas. Deployed async FastAPI services in Docker on a Kubernetes-orchestrated environment with CI/CD.

Built a Graph Neural Network surrogate model to accelerate CFD simulations of heat transfer and cavitation, achieving up to 1000x speedup over conventional solvers. Applied hyperparameter optimization and dimensionality reduction to improve surrogate accuracy and stability; curated simulation datasets with ANSA and ANSYS Fluent, including hydrogen fuel-dispersion studies.