Musings of an AI engineering student pivoting to safety research. Notes on LLM evals, agent harnesses, mech interp, and production lessons.

Madrid. 19. Second-year at Universidad Francisco de Vitoria. Moving from production engineering toward evaluation and alignment research. See about for the long version.

Projects

SAM

closed source

Production LLM agent over a CRM for a Spanish firm that maintains the exteriors of 2,500+ buildings. A declarative profile-pipeline serving Salesforce (sync) and an Android app (SSE streaming).

FastAPIVertex AISalesforceCloud Run

LinceReservations

closed source

Institutional study-room booking platform built to a Madrid university's requirements (pending adoption). Custom OTP + Microsoft SSO, row-level security, five locales, calendar sync.

ReactTypeScriptSupabaseMSAL

Numo

closed source

Local-first desktop statistics app. A real stats engine (pandas / scipy / statsmodels) bundled with the app; your data never leaves your machine, only the interpretation layer talks to a model. In development.

TauriPythonTypeScript

reeljet

open source

AI short-form video ad generator for apps and SaaS, packaged as a Claude Code skill. Real captures plus generated motion, assembled deterministically with ffmpeg.

Claude CodeffmpegHiggsfield

Token Budget Advisor

open source

A Claude Code skill that picks a depth/cost budget before Claude answers, so you don't overpay for simple turns. Merged into the everything-claude-code toolkit.

Claude Code

save-session

open source

A Claude Code skill that writes a compressed, structured summary of a working session to a personal knowledge vault, so context survives across sessions.

Claude Code

More writing