Camille Rousseau
View profile
Internal data analyst chat
Claude agent that translates plain-English questions into SQL against a Postgres warehouse, runs query, returns formatted answer with chart.
Adopted by ops team — 200+ queries/week, no SQL training needed.
0
Problem
Ops team blocked on every metric question — had to ping data eng, wait, get a CSV back, then try to read it. Decisions slowed by days.
Users
Internal — 22-person ops team, finance, founder
Key flow
- 1User asks plain-English question
- 2Agent shows SQL + result + chart, with "explain this answer" toggle
- 3User saves the query if useful, or refines and re-runs
What I built
- Read-only role to the warehouse with row-count and timeout guards
- Claude generates SQL with schema-aware few-shot examples
- Auto-charting via Vega-Lite when result looks time-series
- Saved-query library — non-SQL users can re-run named answers
Limits
- Read-only — never writes, drops, or updates
- Joins capped at 4 tables to keep latency under 5s
AI tools used
- Claude
- Anthropic API
- Supabase
@camille.rousseau
