Writing

Notes from building Thai speech, translation, and document models, written for the people who will call the API.

01August 25, 2026 · 6 min readThai has no word spaces, and it changes how a translation API has to workWhy the absence of spaces between Thai words decides how markup is preserved, how a glossary matches, and where a naive text splitter goes wrong.Where the boundary actually livesMarkup, and why `format` is a fieldA glossary that cannot use word boundariesPruning decides the bill, and its position is fixedChunking, and why structure is the only honest boundaryWhat this means for a caller 02August 25, 2026 · 6 min readCounting the pages in a PDF, without a PDF libraryWhy an OCR endpoint that charges per page has to read the page count itself, and what it takes to do that from the bytes alone.Why this is written by handThe chain from the last byte to the page countFour shapes of the same tableRefusing is the featureA document that lies about its own countImages are simpler, and the trust boundary is the sameWhat the trade actually bought 03August 25, 2026 · 6 min readCharge before inference, refund on failureHow a metered API keeps its books correct when a request is replayed, a stream dies halfway, or the cache it depends on stops answering.The invariant is enforced by the databaseThe stream that dies halfwayThe same request, sent twiceThe cache that stops answeringWhat is left over