Directed development of a production multi-agent LLM system enabling natural-language querying of a live education database.

Context

Alpha Schools needed a way for non-technical staff to query a 1EdTech education database without learning SQL or waiting on an analyst. The bar wasn't a demo — it was a system that had to hold up in daily use, with the access controls and reasoning transparency that implies.

What I Did

• Directed a PM and two business analysts building a production multi-agent LLM system for natural-language querying of the database.
• Specified a Text2SQL architecture using chain-of-thought reasoning and in-context schema learning, so the system could reason about an unfamiliar schema rather than relying on brittle, hard-coded mappings.
• Built role-based access control into the system's core design, not bolted on afterward, given that different staff needed different visibility into student data.
• Set the technical roadmap toward a RAG migration, aimed at cutting latency and improving output determinism as usage scaled.

Capabilities Demonstrated