_{AI Engineering student in Madrid, turning from shipping production systems toward research.
Over the last year and a half I've built production systems: LLM agents over Salesforce taking real traffic, institutional department tooling in real use, a university booking platform built to spec and pending rollout, and a local-first desktop stats app. On my own, I took apart an undocumented binary format and wrote a decompiler for it.
The bugs were the teacher. The bottleneck isn't the engineering, it's the science underneath.
So I'm working my way in from the applied side, with merged fixes to the eval infrastructure behind frontier-model safety and the lessons written down in public.
Builder in the concrete, learner in the deep.}

xabier.me →

Small fixes, real repos — the eval infrastructure behind frontier-model safety research, and the tools around my own work.

• METR/hawk #627 — METR (Model Evaluation & Threat Research). Auth fix in hawk, their cloud Inspect-AI eval runner: a passthrough credential header containing only whitespace is now treated as missing (401 + anonymous) instead of surfacing an unactionable "invalid api key". Merged.
• princeton-pli/hal-harness #182 — HAL, the Holistic Agent Leaderboard (Princeton PLI). Resolved a wandb/weave/gql version conflict that broke every agent-env container at startup; aligned the three runners, the Docker image, packaging, and contributor docs. Merged.
• boykopovar/AnyPS5 #4555 — AnyPS5 (50k stars), a relinker that ports PS5 executables to Linux and Windows. Made libkernel build on macOS: O_DIRECT is now guarded (the macOS SDK doesn't define it) and the getrlimit test gets its missing header. Built and tested on an M3. Merged.
• affaan-m/ECC #920 — everything-claude-code, the most-starred agent toolkit on GitHub. Contributed the Token Budget Advisor skill. Merged.
• karanb192/awesome-claude-code-mods #254 — awesome-claude-code-mods, a catalogue of 6,000+ Claude Code mods. Added the five EMCC mods to its scanner seeds, after fixing what its scanner and the maintainer's review flagged. Merged.

Real systems built for clients and institutions, some live and taking traffic, some built to spec and pending rollout. Code is private; the writeups on SAM and LinceReservations are on xabier.me.

SAM — production LLM agent Commercial + field-ops orchestrator built for a Spanish firm that maintains the exteriors of 2,500+ buildings. A declarative profile-pipeline over Salesforce + Vertex AI / Gemini — adding a new agent is one declarative entry, no core changes. Serves Salesforce Apex (sync) and an Android app (SSE streaming). FastAPI · Cloud Run.	LinceReservations A study-room booking platform I built for a private university in Madrid, designed to their institutional requirements and pending adoption (not yet deployed). Five languages, institutional SSO (Azure AD / MSAL), row-level security, calendar sync. React 19 · TypeScript · Supabase.
Numo A local-first desktop statistics app, in development — a real stats engine (pandas / scipy / statsmodels) bundled with the app, so your data never leaves your machine. Tauri v2.	Institutional tooling Evaluation and automation tools built for a university department — project-review interfaces used by ~30–40 professors a year, room-booking automation, an annual reporting pipeline. Production institutional software, not proofs of concept.

Hubble Decompiler for MultiBase Cosmos: compiled .omd modules back to COOL source that recompiles to the same bytes. The format had no public documentation; it was worked out from the binaries, with an ImHex pattern to read it field by field.	EMCC Five tested Claude Code mods (453 tests): a live panel of what Claude is doing, a context bar, and a guard against moved goalposts. Listed in awesome-claude-code-mods
TBA — Token Budget Advisor Depth/cost choice before Claude answers · merged into ECC	save-session Claude Code skill — compressed session summaries to a vault

_{Also: xabier-blog (this research notebook, Astro, xabier.me) and reeljet (a short-form video ad generator, as a Claude Code skill). Early / idea-stage: harness-ablation (an agent-harness ablation study I sketched out) and xabier-arena-solutions (placeholder for the ARENA curriculum, starting soon).}

• Writing up the lessons as case studies on xabier.me: SAM's agent architecture and the LinceReservations build are up, Hubble's teardown is next.
• Contributing to eval-framework open source — more PRs in the pipeline after the five above.
• Next — the ARENA alignment curriculum, to build the research foundations under the applied experience.

Blog · GitHub · Email · RSS