Lakshman H
building production
GenAI systems.

The client's frontline support team was answering the same three classes of question on repeat — billing and account issues, network troubleshooting, and internal employee policy lookups — each requiring different tools, different knowledge bases, and different escalation paths. A single chatbot couldn't model that.

Designed and built from scratch a multi-agent system on LangGraph — a supervisor agent that classifies the inbound query and routes to one of three specialized agents, each with its own tool set, system prompt, and grounded retrieval over the relevant corner of the knowledge base. Gemini as the underlying LLM. Vertex AI Vector Search indexes thousands of documents — product docs, network runbooks, internal policies — with semantic retrieval. LangSmith captures every trace for failure-mode analysis and iterative tuning. Productionized as a containerized FastAPI service.

The existing recommendation tab was missing context — users got suggestions that looked plausible but had no relationship to current inventory, current pricing, or the specifics of the user's session. A vanilla LLM hallucinated products that didn't exist.

Integrated GPT-4 with a retrieval layer over the live product catalog and the user's session context. The model could only recommend from the retrieved set, with explanations grounded in real attributes. Built prompt templates that proved reusable across other downstream LLM features in the platform.