Selected work
Things I've built
Production systems, not demos. A few are internal, so some repos stay private.
MCP servers for Meta Ads & Google Ads
LiveModel Context Protocol servers that let an LLM query live ad-account data (campaigns, spend, performance) and return structured, ready-to-analyse insights. Read-only and typed, safe to point at production accounts.
Analiza tus arras
LiveReads a Spanish "contrato de arras" and returns its type (penitenciales / confirmatorias / penales), key figures and risks, each grounded in Código Civil citations. A LangGraph agent with two-path RAG over a legal knowledge base, typed Pydantic outputs, an LLM-as-judge eval harness, and a local PII-anonymization layer so no personal data reaches the model. Open source, live on Vercel.
Competitive audit tool
Coming soonScrapes competitor sites and ad libraries, transcribes video creatives with Whisper, and uses Claude to extract structured competitive data (offers, angles, funnels) into a single report.
Weekly Google Workspace digest
Coming soonA weekly automation that gathers everything across Google Workspace (Meet transcripts, Google Chat threads, and Gmail) and uses Claude to produce a single digest of what happened, what's pending, and what needs a reply.
Stack
Tools I work with
A short list I actually reach for, grouped by where it sits.
-
AI / ML
Claude API MCP RAG · pgvector Structured outputs Hermes Agent/Clawbot LangChain (learning) -
Backend
TypeScript Node.js Python Supabase PostgreSQL -
Automation
n8n Google APIs Meta APIs -
Frontend
Next.js Astro Tailwind -
Infra
Vercel Docker Caddy Tailscale
About
Short version
I moved from paid media to building AI systems in about a year, self-taught, by shipping.
My focus is production over prototypes: systems that run every day, fail loudly, and are cheap to keep alive. I care more about a workflow that quietly saves hours than a demo that impresses for five minutes.
Contact
Get in touch
Best over email. I read everything and reply personally.