Agentic Conversational Analytics
Natural-language access to enterprise data — an agent that plans, grounds itself in Snowflake metadata, writes and safely executes SQL, and answers finance and analytics teams in seconds.
- ~30%analyst productivity gain
- NL → SQLself-serve data access
Architecture · simplified
- QuestionPlain-English ask from finance or analytics
- Planner agentLangGraph orchestration
- Metadata groundingSnowflake schema & business context
- Secure SQLGenerated, validated, read-only
- AnswerNumbers, tables and narrative
Context
Finance and analytics teams depended on a small group of data specialists to answer everyday questions. Every new question meant a ticket, a queue, and a handful of SQL written by someone else.
The challenge
Letting people talk to enterprise data is easy to demo and hard to trust. The system had to understand the business’s own vocabulary, produce correct SQL against a real warehouse, and never become a new way to leak or damage data.
What I did
- Architected the platform end-to-end: an agentic workflow built with LangChain and LangGraph that plans a question, gathers context and decides which tools to call.
- Grounded the model in Snowflake metadata, so generated queries use the right tables, columns and business definitions instead of guessing.
- Designed secure SQL execution as a first-class concern — generated queries are validated and run with controlled access.
- Led the team that took it from prototype to a production service used by finance and analytics teams.
Outcome
Analysts got self-serve, natural-language access to their data, improving analyst productivity by ~30% and freeing the data team for deeper work.