At Tata Consultancy Services, I built a production WhatsApp multi-agent AI platform for a leading life insurer, handling policy queries and service requests. It handles more than 135,000 API calls a day across 320+ concurrent sessions.
I improved its RAG pipeline by benchmarking embedding models and selecting Qwen3-Embedding-8B, cutting per-token embedding costs by about 92%. Hybrid keyword and semantic search with reranking reached 93% retrieval accuracy on a 100-query evaluation set.
With LangGraph orchestration, Redis cache lookups, and intent classification, I reached 96% routing accuracy. Retry and backoff middleware and prompt-injection guardrails reduced LLM API failures from 15–20% to near zero.
On QueryDesk, I built a FastAPI orchestrator and rule engine that achieved 97% routing accuracy across 10 categories. I also built AutoHub, a vehicle-data pipeline that discovers brochures, extracts specifications, and runs monthly without manual steps.

