Self-hosted speech recognition and clean-up cluster
The inference backend behind Voicelna, built and operated in-house rather than rented from a speech API. Streaming recognition tuned for sentences that switch language mid-clause, followed by a language-model pass that turns a raw transcript into text worth sending.
Measured behaviour: 144–149 ms server-side from end-of-speech to raw text on a single stream, 8–9 concurrent streams held under 500 ms, and a documented capacity model for scaling past that.
Running our own inference is also what makes the privacy claim real: audio goes to our cluster, not to a third party's API.
AI search visibility, measured on our own site
Being findable by an AI assistant is not the same problem as ranking in a search engine. We built a way to measure it — what the AI crawlers can actually reach, and what they can actually assert — and applied it to this site first.
What that meant in practice: fifteen AI crawlers named explicitly rather than left to infer permission, a structured-data graph describing the organisation and its products, push-based submission so changes are known in hours rather than weeks, and — the part that actually matters — factual content specific enough for a model to cite.
That last point is the whole lesson. Adjectives cannot be cited. Verifiable statements can. Most sites fail here, and no amount of technical work compensates.