Jesse LeVasseur

Collapses the slow, manual grind of public-records requests into one automated pipeline — filing to government agencies and turning what comes back into structured, searchable data.

• Built to fail closed, on purpose. I knew the AI would invent answers if I let it, so I required it to check every answer against its source, do the dollar math in code, and route anything uncertain to a human before it goes out. The safety design is the point.
• Reads scanned government records and pulls the data — including how each board member voted — as exact quotes from the page, so it can't report a result that isn't there.

Runs a membership organization's real parliamentary process in software — motions, debate, roll-call votes, minutes, and bylaw amendments.

• Reused my first platform's design to stand the whole thing up in two weeks: clause-anchored comments and redline on the governing documents, a 2/3-threshold voting engine with eligibility rules, and an append-only public audit log.
• It needed sign-in flows no off-the-shelf tool covered, so it was built to the OAuth 2.1 spec to power an AI (MCP) connector.

A free, public census of ~8,000 school-district library catalogs. It maps the undocumented internal APIs behind six different catalog systems and checks them at national scale — every result cited to its own source.

• When it started bleeding ~$160/month in cloud costs, I caught it and had the analysis moved onto a local copy — same answers, no bill.
• An AI review pipeline reads each title on a mix of local and cloud models at about $0.09 a record, and every tag has to be quotable from its source or it's dropped.