AI & Data Engineer · Santiago, Chile · Google Cloud Generative AI Leader
I work at the point where data engineering meets applied AI. Ten years building production data pipelines on GCP, and the last two shipping RAG systems that are measured rather than demoed.
The thing I care about: an AI system you can't evaluate is a prototype, whatever it is running on. Retrieval quality belongs in CI next to the unit tests.
That applies to how I build, too. I lean on AI heavily and measure everything it
produces — the DECISIONS.md in the RAG repo below is what the measuring part
looks like, and what the tooling can't hand you.
Industrial RAG in production — On-premise system for a manufacturing client in Spain. Operators query machine manuals, maintenance procedures and captured plant knowledge from tablets on the floor.
- ~16,000 queries in a 7-day window
- RAGAS faithfulness 0.96 median, context precision 0.997, captured weekly from production and versioned as baselines
- 307 automated tests and 5 CI pipelines — tests, security, regression, retrieval eval, baseline capture
- 10 containers: FastAPI + SQLAlchemy async, PostgreSQL, Qdrant, Ollama, Celery + Redis, Vue 3 PWA, nginx, NVIDIA GPU
Analytical platform for a national retail group — 2,702 Dataform models in production across seven repositories, layered STG → DW → PUB: conformed dimensions, fact tables, pricing and campaign domains.
Tax and banking data extraction — Multi-tenant service that connects directly to Chile's tax authority and six bank portals via Playwright. 45 FastAPI routers, 162 test files, swappable adapter behind one environment variable so the downstream contract never changes.
Real-time voice agent — Twilio → Deepgram → Groq → Cartesia, bidirectional mulaw/8000 audio over WebSocket with a jitter buffer. Spanish-language collections and customer service.
Extracted from production systems — the transferable parts, with the hard decisions documented:
rag-hybrid-citations
Hybrid RAG on Postgres + pgvector. Dense and BM25 fused with Reciprocal Rank
Fusion inside SQL, citations that carry document, page and section so a reader
can verify them, and an explicit refusal path for when the corpus doesn't hold
the answer. Its DECISIONS.md covers eight bugs paid for in production —
including three chained failures where each fix caused the next.
guided-visual-check Reference-guided visual inspection. The model reports evidence with a confidence; the decision to report a failure or send it to a human lives in twenty lines of auditable code, not in the prompt. Angles and orientations are handed over measured, because that is the question vision models answer confidently and wrongly.
sistema-helper-en Technical interview trainer in English, built on a real-time voice pipeline. The numbers are computed by code and the judgement comes from the model, shown separately — if the metric you measure progress with can hallucinate, the progress you see is noise.
surveybq-engine CSAT/NPS survey engine with no dependencies and no build step. Responder, visual editor, analytics panel, BigQuery output. Static files, because whoever scans a QR code arrives on a clean browser, on some phone, sometimes without signal — every kilobyte and every build step is a way for that to fail.
Reference implementations of patterns I use, written as synthetic code so the pattern is visible without any client material: gcp-etl-pipeline — Beam on Dataflow, Pub/Sub, Dataform, Terraform · rag-crag-reference — CRAG dual-judge grading, reranker circuit breaker, RAGAS eval · multi-tenant-saas-starter — tenant isolation from day one · case studies — five production systems, how each was decided and measured
AI RAG in production · evaluation with RAGAS · hallucination detection · reranking with circuit breakers · Qdrant · pgvector + HNSW · OpenAI · Claude · Gemini · Mistral · Ollama · faster-whisper
Data BigQuery · Dataform · DBT · Apache Beam · Dataflow · Pub/Sub · Airflow · Polars · PostgreSQL
Backend Python · FastAPI · SQLAlchemy 2.0 async · Celery · TypeScript · Next.js · tRPC
Ops Docker · on-premise GPU deployment · nginx · Caddy · GCP · Vercel · GitHub Actions · Prometheus · OpenTelemetry · Sentry
Open to remote roles. Spanish native; English is professional in reading and writing, conversational spoken.