AI Lab
Research, experiments and notes from the lab.
Working notes from live builds. Written for engineers and operators, and revised whenever a build proves them wrong.
Most disappointing RAG systems are not let down by the embedding model. They are let down by chunks that split a sentence in half, or that bury the one relevant clause inside a thousand words of boilerplate. Fix boundaries and overlap before reaching for a better model.
Every demo in this lab runs with no API key: browser speech synthesis, browser speech recognition and keyword retrieval carry the whole experience. Paid models are an upgrade, not a dependency. A feature that cannot degrade is a feature that cannot be demoed.
A caller forgives a slightly wrong answer and hangs up on a two second pause. The budget that matters is the round trip: capture, transcribe, decide, synthesise, play. Everything else is decoration.