Conserje
A chat that qualifies leads, routed through n8n
The chat on this site. A short conversation instead of a form: it extracts a brief, scores it, and hands it to n8n workflows for notifications, email and CRM.
Tools I use every day, and labs where I take the machinery apart to understand it better and get closer to what I want to be able to offer.
A chat that qualifies leads, routed through n8n
The chat on this site. A short conversation instead of a form: it extracts a brief, scores it, and hands it to n8n workflows for notifications, email and CRM.
Project memory for AI coding assistants
An MCP server that gives Claude Code and Cursor the context of each project: decisions, conventions and the history that never lives in the code. Hybrid search, strict isolation between projects.
A music studio where no stage advances without the artist
No agent can mark a stage approved, and the tests fuzz every possible action to prove it. Every file records which engine produced it, and the release disclosure is computed from that log.
AI agents that ask before they act
Agents declare every action first. Low-risk actions run; anything touching money, customers or personal data waits for a named person.
Watches public tenders so a person decides which to bid for
Reads Colombia’s open procurement data, flags what is new since the last run and matches it against a company’s profile. Every reason it gives is checked against the record. It recommends; it never bids.
Which parts of a Salesforce release touch your org
Reads the release notes, maps an org’s metadata and returns what may break and what to test, each point quoted from the source. The tool does the reading; the consultant decides.
Where I measure agents and learn the layers underneath: tokenizers, sampling and, next, training.
Does your agent ask before it acts?
An eval harness that measures when a tool-using agent acts, asks a person or refuses, including when a document tries to hijack it. The score comes from what the agent did, not from what it said.
Lab notes on what happens inside a language model
A tokenizer and samplers written from scratch, run on a small local model. First finding: the same text costs 7–42% more tokens in Spanish than in English.
Training small models, from scratch to fine-tuning
The labs where I train instead of prompt: a transformer from scratch, LoRA, preference tuning and reinforcement learning with verifiable rewards. Planned and budgeted, not run yet.
A writing editor where the model only proposes
Every request carries your canon and the passages from the rest of your work that bear on it, so an answer never arrives context-free. Self-hosted and single-user.
Reader components for digital stories
One text, five reading modes: scroll, book, terminal, editorial and hypertext.