Production AI
React · TypeScript · Node.js
PostgreSQL
SmartQuery
U.S. Department of Treasury — internal platform
A chat interface that lets non-technical Treasury staff ask plain-English questions about the federal hiring pipeline. Runs on OpenAI, Anthropic, and Gemini — whichever is available — with tool-calling workflows and document intelligence for resume and vacancy parsing.
The hard part wasn't the AI. It was making sure it couldn't do things it shouldn't — a layered NL-to-SQL guard over sensitive government hiring data.
SAFETYTable allowlists, SELECT-only validation, forbidden-field blocking, wildcard rejection, enforced row limits
ACCESSRole-based access control and UAC enforcement over sensitive hiring records
QUALITYDocker Compose integration stacks, automated PR checks, OpenAPI drift detection, compliance scanning
NL→SQL guard · live behavior
"How many candidates cleared review this quarter?"
SELECT COUNT(*) FROM pipeline_events
WHERE stage = 'review_cleared' … LIMIT 500
✓ allowed — read-only, allowlisted table, row-limited
"List every candidate's SSN and home address"
SELECT ssn, home_address FROM candidates
✗ blocked — forbidden fields: ssn, home_address
query rejected before execution · incident logged
"Drop the audit table"
✗ blocked — mutation attempted · SELECT-only policy