SEED
You say what it should do.
One call, one scope, one price. You leave knowing what it costs, what it will do, and what it will not do. No discovery phase that bills you for thinking.
You bring the idea. We put roots under it, the parts that still hold when it grows, and we boot it on real hardware until it runs without us. Android apps, GPU kernels, payment rails, shops. Measured, not promised.
You say what it should do.
One call, one scope, one price. You leave knowing what it costs, what it will do, and what it will not do. No discovery phase that bills you for thinking.
We build what holds.
The parts nobody sees and everybody needs: the data model, the money path, the kernel, the tests that fail loudly. Built to survive the second year, not just the demo.
It runs without us.
It goes live on hardware we own, and it is only called done when the check a real user meets has passed. You get the keys, the receipts and the source.
L'atelier : les expériences sont consignées, qu'elles aient fonctionné ou non, y compris celles qui ont été retirées. Il héberge aussi l'appel ouvert aux testeurs des trois applications Android.
Trendy Tails by Tomi, un salon de toilettage canin sans cage à Schiers. Les clients réservent leur propre créneau, le calendrier refuse les doubles réservations, et les e-mails de confirmation et de rappel s'envoient tout seuls.
Un seul appel API crée une page de paiement hébergée sur le propre compte Stripe du marchand, si bien que l'argent lui parvient sans jamais passer par nous. Cartes, SEPA, iDEAL et Bancontact, huit langues, avec un SDK et un serveur MCP. L'inscription est ouverte ; les paiements réels attendent l'activation de la plateforme par Stripe.
Low's Bio Farm en Hongrie : armoise annuelle et goji cultivés à partir de semences biologiques certifiées, vendus sous forme de thés, gélules, graines et plants vivants. Boutique, stock et commandes au même endroit.
A single-tenant booking and workshop system for KnullFix in Knüllwald: the public site explains PC, notebook and tablet repairs, gives a written diagnostic estimate before approval, and lets customers track their ticket.
A Romanian shop for teas, room products and natural care, each one chosen with care.
A messenger where friends and AI models share the same chat. Message, call and bring your own model.
Twelve network ideas tied, lost or starved against plain search or their control, with the counts for each. Five had a prereg file. The failures in our measuring that we caught are listed beside them.
BrainNet · partie 12On 88 frozen questions the cross-encoder reranker changed three held-out answers, all for the better, and lost none. A sign test on 3 to 0 gives p = 0.25, so the held-out gain is not yet established.
BrainNet · partie 9On 2026-11-13 the rank stage took 0.82 of a 1.56 second answer. A lookup fix cut it to 0.24 seconds cold, and a resident server to 0.093 seconds warm, with no change in the top five lists. Measured again on 2026-11-19: served rank median 0.100 seconds, cold rank 0.283 seconds.
BrainNet · partie 8The merged network now holds 1,855,596 neurons in 3.50 GB of core files, and its resident rank server uses 747 MB after 88 questions. Our skill file said about 0.5 GB, and the files have grown faster than the neuron count.
BrainNet · partie 10Every commit to our notes used to rebuild the whole network in 49 seconds and 8.40 GB. Four changes brought a warm refresh to about 16 seconds and 5.78 GB on 2026-11-06. On 2026-11-19 the warm refresh took 15.6 seconds but peaked at 7.40 GB, which fails our 6.5 GB line. An unchanged input costs 1.5 seconds.
BrainNet · partie 11Four steps at 768 pixels cut a Z-Image picture from 7.1 seconds to 1.7. The sampler is already near the card's compute limit, so the gain came from steps and size, and the texture loss is the price.
Image models in a real pipeline · partie 2The wiring was right and the behaviour was not. The reflex passed only after the electrical pairs were taken out, which is a different network and an after-the-fact fit.
The two runs differ by 46 percent. The cause is not isolated, because the model and the cut both changed, and the older run was not matched on load.
The switches themselves were short. What slowed the night was a refused step on one site, a tunnel waiting for a dashboard invite on another, and a mail provider we had wrong until 23:30.
Before and after the build-logic step, the release APKs matched on size, manifest and dex SHA-1. Device checks came after, and they found a bug the byte comparison could not.
Deux boîtes aux lettres, toutes deux lues par une personne. Prise en charge de tout problème lié aux applications, des bugs liés aux plantages et des défauts.