I'm Xabi, 2nd-year AI Systems Engineering at UFV (Madrid). I've spent the last year and a half shipping production agent systems — Salesforce-integrated commercial agents, MCP servers with security review iterations, field-ops trainers on Android, multi-language institutional tools. Now pivoting toward research output as the prerequisite for the work I actually want to do.
- Writing — empirical ablation study of LLM agent harness components (harness-ablation)
- Studying — ARENA curriculum, chapter 0 fundamentals (solutions repo)
- Contributing — merged fixes to the METR/hawk and HAL eval harnesses
- Reading — Anthropic engineering posts on eval methodology + infrastructure noise
- Blog — in setup; first posts on the journey + paper reading notes
|
Empirical ablation study of LLM agent harness components — prompt structure, tool selection logic, retry policy, context management, error handling. Extends Anthropic's Quantifying infrastructure noise in agentic coding evals outward: if infra config dominates eval variance, what does the rest of the harness do? Synthetic controlled experiment informed by production deployment experience.
Status: paper in progress |
Public journey through ARENA — Alignment Research Engineer Accelerator curriculum — plus the original mech interp experiments I run alongside each chapter. Builder background, weak research foundations, trying to close the gap before applying to alignment/interp work.
Status: chapter 0 in progress |
Merged contributions — fixes to the eval infrastructure behind frontier-model safety research.
- 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/gqlversion conflict that broke every agent-env container at startup; aligned the three runners, the Docker image, packaging, and contributor docs. Merged. - affaan-m/ECC #920 — everything-claude-code. Contributed the Token Budget Advisor skill. Merged.
|
Full-stack room booking system deployed at UFV's DOT space. 12 rooms, 4 user roles (student, professor, PAS, admin), real-time sync via Supabase Realtime, Outlook Calendar integration via Microsoft Graph, institutional email via Office 365, admin dashboard with audit logs, blacklist system, XLSX report exports, and 5-language i18n.
|
Claude Code skill that intercepts your prompt, estimates token consumption with a zero-dependency heuristic engine (measured ~87.6% accuracy vs Contributed to everything-claude-code (ECC) by affaan-m — the agent harness with 220k+ stars used across the Claude Code ecosystem.
|
|
Claude Code skill that saves compressed session summaries to an Obsidian vault, automatically maintaining bidirectional links between project notes, session logs, and technology nodes.
|
Portfolio of internal web tools for UFV's DOT space — operational management interfaces built as lightweight applications deployed on Vercel.
|
Languages
Frameworks & Tooling
Platform & Data
AI & Agents
Merged into the eval infrastructure behind frontier-model safety research
Focus — agent evals · Inspect-AI · mechanistic interpretability
Working in — Python · TypeScript · PyTorch
