Andrii AI

The Apostille Handbook is a live guide to legalising documents for use abroad, where every fact carries its quote, source and date. Try it at https://andrii.ai/apostille. The problem: people send documents overseas once every few years, so they shouldn't have to study government sites that change quietly. Washington's apostille counter left Olympia for Tumwater in June 2025, yet searches still surface superseded forms. On 13 October 2026 even our own site sent Viet Nam customers to consular legalization, a month after the Apostille Convention took effect there. Andrii Co. is a Washington notary practice, so we built this for a problem we meet firsthand. How it works: the system watches official pages. When one changes, Muse Spark drafts proposed facts from the added lines only. Deterministic code then checks that each quote appears verbatim on the captured page and that the value sits inside its quote. A person approves every fact in a review queue, and approvals reach the public page live over WebSocket. Nothing is public until a person accepts it. Staff can also quote an exact span of a page when the model misses something. Live today: an index of 411 authorities across 133 countries, guides with citations, a "what changed" panel per source, passkey sign-in, a photo precheck, order drafts, and staff tools for review, tracing and job queues. A new country is new data, not new code. Measured: on a 30-case gold set, run live on 2026-11-02 and producing 68 drafted records, 100% of quotes were found verbatim, 98.5% of values appeared inside their quote, and 63.2% of records were filed under the expected topic (this understates, because each case has one topic but drafts several records). The verifier caught 181 of 182 planted faults. The gold set is synthetic passages, not live government pages. Built to hold up in production: a single Rust binary with SQLite, additive-only schema changes, nightly backups with a restore check, per-IP rate limits and cost caps on public AI surfaces, and a scoreboard script that checks each deploy before and after, rolling back on regression. Every Meta API call and downloaded page is recorded and replayable. Not finished: only 2 of the 411 authorities have approved facts so far (24 facts), and the rest are indexed and waiting for review. No customer has been charged, demo document photos are synthetic, the email inbox is off in production, and the first live FedEx label is still to come. Judges can run it with no keys: ops/judge/preflight.sh && cargo xtask judge. It replays recorded Muse Spark answers and fakes every outside service.