Agent with tools, not a raw LLM
The model never guesses numbers. It calls SQL tools against the real database, so every answer is grounded in actual data instead of the model's imagination.
Project Overview · Agentic AI + Data
UCSD Smart Analyst
An AI agent over 270,000+ volleyball points that answers coaches with real, validated stats and visualizations — the team-specific analysis general-purpose LLMs can't do.
- 270K+ Volleyball points made queryable
- NCAA D1 Built for the UCSD coaching staff
- Validated Real SQL, not hallucinated stats
The problem
Coaches need answers their data can prove.
UCSD's volleyball coaches wanted to ask real questions of their data — how an opponent sides out against a particular serve, how a lineup shifts win probability — but the numbers live in 270,000+ raw VolleyMetrics points. A normal dashboard can't answer open-ended questions, and a general-purpose LLM will confidently make up plausible-sounding stats.
Coaches needed something that could reason over the actual data and be trusted not to invent a number.
Architecture
An agent that queries the database, then checks its own work.
A coach asks in plain language. An OpenAI agent introspects the Postgres schema, writes a SQL query, runs it against Neon, and validates the result is correct before answering — retrying if the query is wrong. Only verified numbers reach the coach, rendered as stats and Plotly charts.
Key decisions
Why it's built this way.
A correctness feedback loop
Each query is validated before results surface. An invalid query triggers a retry, not a wrong number handed to a coach who's about to make a decision on it.
Schema introspection
The agent reads the live schema, so it writes queries that match the actual tables and columns rather than hallucinating a field that doesn't exist.
Neon Postgres over 270k+ points
The agent collected, cleaned, and loaded VolleyMetrics data into a queryable relational store, turning a messy export into something you can actually ask questions of.
What shipped
Coaches ask; it answers with sourced stats.
Coaches ask questions in plain language and get back real, sourced stats and visualizations that general-purpose tools can't produce. Here's a walkthrough.
Challenges & what's next
Honest edges.
The hard part was trust: making the agent reliably produce correct SQL and catch its own mistakes before a coach ever saw a wrong number. The validation loop was the key — it's the difference between a demo and something coaches will actually rely on. Next: caching common queries for speed and a polished coach-facing chat UI.
Stack