The same sentence, read by the machine as it learns — scroll through the run
STEP 6,000
fetching earlier selves of the model…
In this edition
Custom AI systems
Retrieval pipelines over your documents, model evaluation and fine-tuning, language-model
features built into your product. The reference build runs in production: 1.3 million
chunks, answers with receipts.
No page builders, no templates: sites written by hand, fast and private by construction —
zero trackers as a contract term. Forty-plus pages run this way in production, in university
classrooms, every term.
Scraping and data curation, document processing, programmatic publishing. The corpus behind
the billion-parameter model — 257,918 articles — was collected by pipelines like the ones
I build for clients.
PULLMAN, WA — four university courses, every termWASHINGTON STATE — Global South Solidarities, registered nonprofitNEW YORK — USD 2,000 Modal compute grant
REMOTE-FIRST · IST +5:30 · REPLIES WITHIN 24H
Selected results — abridged
Figure
Claim
Evidence
109 params
Language model designed and trained entirely from scratch, on consumer hardware plus free cloud quota. In beta.
Four complete pages in registers clients actually ask for — dark aurora,
loud pop, calm pastel, silent monochrome — fully animated, nothing like this page on purpose.