Retrieval for tax agents
State tax law, crawled, chunked and embedded so AI agents used by tens of thousands of tax professionals can cite the right authority.
- Role
- Software Engineer
- When
- Aug 2026 – Now
- Where
- San Francisco
- Stack
state tax jurisdictions covered
faster crawling per search
tools in an MCP server in the Claude directory
pull requests merged in 13 weeks
The problem
A tax answer is only useful if it points at the right authority. Federal sources were well covered; state law was not. Each state publishes statutes, regulations and guidance in its own format, on its own site, and the agents answering state questions had little authoritative text to retrieve from.
What I built
- Ingestion and retrieval for 11 states. Crawlers, chunking and embeddings that put each jurisdiction’s primary sources into a vector store the agents query. On our evaluations, both citation quality and answer accuracy improved substantially.
- One framework instead of eleven pipelines. The first states were built one at a time. I folded them into a single parameterized framework in 11 days: adding a state is now a module and a config block, and the code covers more states in 14% fewer lines.
- Faster live search. Replaced batch scrape jobs with parallel per-URL fetches, cutting crawl time per search by over 90%.
- An MCP server in the Claude directory. Took a 7-tool server from handover to approval, implementing OAuth 2.0, streaming HTTP, scopes and metering.
How I work there
In my first 13 weeks I merged 134 pull requests and reviewed 324, across React and TypeScript on the front end and Django, Flask and PostgreSQL behind it. I also carry production reliability for the pieces I own, including alerting on tool and session errors.