AI & Automation Engineer — I build AI-powered systems for small businesses at Luzel Estudio (Zaragoza, Spain): websites, Shopify stores, CRM + AI automation, and the workflows that glue them together.
🎓 Microsoft Certified: Azure AI Fundamentals (AI-900) · Power Platform Fundamentals (PL-900) — currently on the path to PL-200 and AI-103 (Azure AI Engineer).
Python · FastAPI · Pydantic · pandas · Claude / Anthropic API · n8n · Power Platform · Azure AI · MCP · SQLite · GitHub Actions
Each repo is one level of the path from workflow automation to production AI engineering — all typed (mypy strict), tested, CI-gated and documented with honest limitations. The reliability kit audits the others in CI:
| repo | what it shows |
|---|---|
| agent-reliability-kit ⭐ | Observability + evals + CI quality gate for LLM agents — the market's #1 blocker (agent quality) turned into code. Span tracing (cost/tokens), golden datasets that reject suites without planted failures, LLM-as-judge with a versioned rubric and a public fake, baseline regression detection that breaks the build. It audits the repos below in CI. 49 tests, 98% coverage. |
| llm-gateway | Cost-aware LLM gateway: declarative capability routing (cost/latency/quality), per-client budgets checked before the call, provider failover driven by an error taxonomy, EN/ES prompt-injection + Spanish-PII guardrails, Prometheus metrics — zero deps in core, 50 tests, 99% coverage. |
| durable-agent-runtime | Durable execution for agent workflows: checkpoint after every step, resume-after-crash without re-running completed steps, human-in-the-loop interrupts, time-travel (fork), and pluggable stores certified by a shared conformance suite. 42 tests, 100% coverage. |
| agentic-orchestrator | Multi-agent engine built from first principles: planner→specialist→critic loop, structural human-in-the-loop approval gate, JSONL traces, parallel DAG execution with tested sequential parity, eval harness with a safety invariant, and a hand-written MCP server. Audited by the reliability kit. |
| rag-knowledge-assistant | RAG where retrieval is measured, not assumed: TF-IDF vs BM25 vs hybrid-RRF compared with hit@k/recall/MRR on a paraphrase-hardened eval set, refusal policy tuned from data, Azure AI Search adapter, eval gate in CI, transparency card. Audited by the reliability kit. |
| ai-client-intake | Production LLM pattern: untrusted model output → strict Pydantic contracts → never-drop-a-lead failure policy → audit trail. FastAPI + Docker + zero-cost mock mode. |
| smb-finance-toolkit | Spanish VAT (Modelo 303) aggregation CLI: integer-cents arithmetic, per-row rejection reporting, domain rules in the model. |
| n8n-to-power-automate | Real production workflow migrations, documented node-by-node: connector mapping, expression translation, and what got worse, not just better. |
All of it comes from real problems at Luzel Estudio — the demos use fictional data, the patterns run in production.
- 📚 Studying PL-200 (Power Platform Functional Consultant)
- 🐍 Deepening Python + Azure AI (AI-103 roadmap)
- 🤖 Building AI-assisted client & finance systems for SMBs