Data · Internal tooling
An AI SQL assistant that non-analysts trust
Node.jsOpenAI APIMCPRedisPostgreSQL
Problem
Non-technical stakeholders filed data requests via ticket. Analysts were spending roughly 30% of their week writing repetitive ad-hoc SQL, and turnaround times for basic questions ran into days.
Solution
Built an MCP-backed assistant that grounds LLM query generation in the actual warehouse schema, caches semantic mappings in Redis, and returns explained SQL with a run / edit / re-ask surface. Every generated query is logged and reviewable.
Impact
62% of non-technical queries are answered without an analyst touching them. Analyst time on ad-hoc SQL fell from ~12 hours a week to under 4. Query correctness, hand-audited on a sample, held at 91%.
Measured impact
62%
Self-serve rate
8h / wk
Analyst time saved
91%
SQL correctness (audit)
Have a system that's failing under real load?
That's the work we take on. Send a short brief and we'll tell you honestly whether it's a fit.
Next · CS · 01
Streaming vulnerability report generation