Notes
Checked, not claimed.
What came up while working, and how much of it survives checking: what holds up, what does not, and what it actually costs. No announcements, no trend pieces.
How a machine finds the right paragraph
Before a model can answer a question about your documents, something has to decide which three paragraphs it gets to read. That choice is the whole game — and you can watch it happen, on a real contract, in the figure below.
When the answer is not in a paragraph
Similarity search cannot answer "what breaks if I change this". A graph can. I ran graphify over the source of this website — 105 files in 2.2 seconds, no model involved — and checked every edge it guessed.
Teaching a local model not to make things up
A model that invents plausible-sounding answers is worse than one that stays silent. How to shut down hallucination in a local RAG pipeline without the model going quiet — measured across 74 questions, not asserted.
What bad audio costs a speech model
Clients don't record in a studio. They call from mobiles, dictate in vans, and sit at the far end of a meeting table. A small experiment on how much each of those actually costs a local transcription model — and where it finally breaks.
Open models in practice
What open models can do, what hardware they need, and which licences actually permit commercial use.
What a local model actually costs
"Buy a server instead of paying for an API" is too short a calculation. The cost blocks that genuinely apply when you run models yourself, and the questions worth settling before you buy anything.