Writes

DeepSeek-V4.1-Flash doubles its backbone to 552B parameters yet holds global KV cache to 890 bytes per token, because the decoder reads a cache the encoder already built instead of deriving its own.

OnlinePlumber captures leads from a Livewire web portal, an Android app, and bulk CSV import - three channels that stayed on two different Laravel auth guards on purpose, funneled into one inquiry status pipeline.

Invobot's webhook, admin, and agent-console endpoints all needed the same signature check, payload parse, conversation lookup, and rate limit - copying them per handler was how a fix in one place stopped reaching the others, so they became composed FastAPI dependencies instead.

Invobot collapses four independent booleans that could disagree into one authoritative status field, checked at the moment a reply is sent, so a bot and a human agent can never both answer the same message.

StudySmart AI split student and admin access into separate API contracts before either interface existed, so a Laravel dashboard added months later needed zero renegotiation with the schema already in use.

StudySmart AI's quiz generation runs behind a strict typed contract so three different clients never have to guard against what the model actually returned.

Fitting a TF-IDF vectorizer on just the pair being compared caps the weight spread between a shared word and a distinctive one at 1.4x, opening the door to false near-duplicate matches.

Why filemetric-engine's cache uses one shared SQLite connection behind a lock instead of one per worker thread, and why WAL mode doesn't make that lock unnecessary.

Why filemetric-engine batches index rebuilds behind a pending buffer so uploads stay an O(1) append instead of a full refit.

A path-keyed cache reprocesses the same file under every filename it's uploaded as - filemetric-engine keys on a SHA-256 of the bytes instead.

Why a prebuilt TF-IDF index beats fitting a fresh vectorizer per file pair - the same corpus doesn't need rebuilding on every request.

Why StudySmart's file comparison was slow, and the index, cache, and write buffer in filemetric-engine that fixed it.