A generative analytics assistant over vendor trading data

The problem

Vendor orders, budgets, invoices and monthly transactions sat in Snowflake at a scale only the data team could navigate. Every commercial question became a ticket, and the queue was the bottleneck on decisions.

What we built

1. Built a multi-LLM pipeline with LangChain that converts a natural-language question into accurate, optimised SQL, then summarises the result set in plain language.
2. Handled complex, large-scale structured data spanning vendor orders, budgets, invoices and monthly transactions, with real-time querying against Snowflake.
3. Combined GPT-4o and LLaMA 70B on Groq, routing between them for cost-effective coverage across simple and complex queries.
4. Delivered it as a responsive chat product with a React front end on a FastAPI backend.