WritingNotes from the build

Working things out in public

Written while building, not after. Mostly about how generative models actually sample, why retrieval quality decides everything downstream, and what breaks in production.

Published 1

EssayGenerative models9 min

Understanding generative AI

VAEs, GANs and diffusion models are three answers to one question: how do you sample from a distribution nobody ever wrote down? Each makes a different bargain between sample quality, training stability, and how much of the distribution you actually cover.

Read →

Drafting 2

NoteInfrastructure

FastAPI on Azure, honestly

What a Python inference service costs to run once the traffic is real: cold starts, worker counts, and the point where serverless stops being the cheap option.

In progress

NoteRetrieval

Chunking is the whole ballgame

Retrieval failures usually get blamed on the model. Most of the time the answer was never in the retrieved context, because the document was cut in the wrong place.

In progress

Elsewhere

Some ideas explain better out loud than on a page. Those are heading for the video shelf — scripts are written, nothing is recorded yet.

See what is coming