Civic data agents
Government data, answered in ten seconds
- Year
- 2024
- Client
- KiwiTech
- Domain
- GovTech
- My role
- Senior ML Engineer · Tech Lead
- Stack
2m → 10squery response
65% → 91%answer accuracy
10K+daily users
The problem
Non-profits and policy researchers needed answers that span voter registration, legislation and subsidy data — spread across slow, inconsistent government APIs.
What I built
- Agents that orchestrate the government APIs asynchronously instead of one call at a time.
- A query classifier that sends simple questions to cache and complex ones to the agents.
- Semantic search across sources with an API layer handling 100,000+ requests a day at 99.9% uptime.
The hard part
Most questions didn’t need an LLM at all. Routing them away from the agents made the whole system both faster and more accurate.
Outcome
Responses went from 2 minutes to 10 seconds, accuracy rose from 65% to 91% and satisfaction from 3.2 to 4.7/5.
Client data and code are confidential; figures are from the delivered system.
Next caseRegulatory document QC →