Skip to content
ibruno

Castellón, Spain · AI infrastructure

We do AI. Starting with the model.

A language model of our own, trained from scratch; an agent runtime that runs inside the company; and applications with the AI built into the product. Two people in Castellón and nine builds, each one explained in depth on its own sheet.

A model that chooses whom to ask

QSE · similarity routing in NAR · Example · sample data

Each step is a token: its vector is compared with the 24 signatures and only the two closest answer.

Mathematics · iTemple

Seven new bounds in Brouwer's table

The iTemple laboratory has improved seven lower bounds for binary constant weight codes, an open problem in combinatorics. Two are already published, with a DOI and a verifier anyone can run.

Catalogue

Nine builds

Each one has its own sheet: what it is, who it is for, how it works and its figures. Nine ways of taking AI into real work, from the model to the application.

iForjaAn agent runtime that installs inside the company. Your content never passes through our servers: it goes from your machine to your AI provider, or it does not leave at all if you use a local model.Agent runtimeiforja.comiCrisolA post-Transformer architecture that keeps living memory across sessions, carries causality inside the forward pass, and imports knowledge the way you install an app. And it is already trained from scratch.Post-Transformer architectureicrisol.comLideriaIt walks a company from commercial prospecting through to compliance with the European AI Act. Eleven agents, one case file per project, and an implementation plan that writes itself. All inside the house: local model, local index, zero cloud.Sovereign AI platform over MCPwww.ibruno.es/lideriaMEENot a database, and not an LLM. A mind of linked scenes that learns by writing, in which the language model holds no authority.Symbolic cognitive organismwww.ibruno.es/meeiTitanProThe S&P 500 analysis engine, in its professional edition.S&P 500 analysisititanpro.comMiAnalistaThe same analysis engine, in an app anyone can follow.S&P 500 analysismianalista.comiTempleA laboratory that searches for architectures, algorithms and mathematical constructions on its own, and keeps every experiment so it can be repeated. Its first result: seven new bounds in Brouwer's table of constant weight codes.Search laboratory · as in tempering steelwww.ibruno.es/itempleiPreparaTakes someone who doesn't code from “I want to predict this” to a trained model, explained and serving predictions, in Spanish, without the data ever leaving their computer.Guided machine learningwww.ibruno.es/ipreparaiEscribanoA writing studio for long-form fiction: it fixes a Bible of the book that enters every generation, and runs six deterministic checks after each chapter, so that by chapter 30 nobody has changed name or come back from the dead.Local writing studiowww.ibruno.es/iescribano

What we do

From architecture to product

Our own language models

Architecture design, corpus preparation and training on our own GPUs. We start from an architecture of our own, from the very first step.

Governed agents

Agents that work on real company documents and data under explicit permissions, with every action they take audited.

AI embedded in the product

AI inside the workflow, not bolted onto it: analysing, drafting, classifying and deciding where the person already works.

Sovereign deployment

All of the above running inside the client's own infrastructure, with their key and their data; with a local model, never leaving for anyone else's cloud.

Read the technical paper

Team

Father and son, one workshop

One has watched enough things break to know what holds. The other trained inside AI, not before it existed. Both are needed to build this, and they are almost never in the same person.

Photograph of Bruno Martí

Architecture and engineering

Bruno Martí

He handles the architecture and the engineering: the post-Transformer architecture of iCrisol, trained from scratch on in-house GPUs; the iForja runtime, about 570,000 lines of code and nearly 3,700 commits; MEE's Rust ledger; the iTemple laboratory, which has improved seven bounds in Brouwer's table of constant weight codes (the first two published on Zenodo under his name); the Lideria platform. More than twenty thousand automated tests across them. Very few people train a language model from scratch, and that is where the difference shows between having read about an architecture and having made one converge. He is also the one still there when a project stops being the interesting part and becomes the part that has to be finished.

LinkedIn
Photograph of Daniel Martí

Product and go-to-market

Daniel Martí

Industrial Design and Product Development engineer (UJI). Digitised a pipe-prefabrication factory end to end: replaced paper with a production-management application he built himself (TypeScript, React, Next.js, PostgreSQL), used daily by every department. He owns the product and how it reaches the market: what gets said, through which channels, and to whom. He trained inside this field, so he has worked with today's tools from the start.

LinkedIn

Contact

A question, an email

For any question about these builds, write to us. A person answers, not a form.

Write to us at

bruno@ibruno.es

LinkedIn · BrunoLinkedIn · Daniel