Software engineer, backend and AI. Gijón, Spain. EU citizen, open to relocation.
By day I work on a warehouse management platform that runs live logistics sites, mostly on transactional workflows and the data access layer (C#, .NET, SQL Server). Before that I spent an internship building tooling to evaluate language models and red-teaming them for information leakage, unsafe output and tool misuse, and I've done computer-vision model testing for IKEA and William Hill. Now I also build the agents.
I contribute to the AI and scientific Python projects I use:
- google/adk-python: fixes in evaluation and memory. Multi-run scoring now treats the efficiency metrics (
token_usage_v1and the rest) as lower-is-better (#8166), andInMemoryMemoryServicesearch handles Thai and Hindi text correctly (#8169). - cupy/cupy: NumPy and SciPy API compatibility: default arguments for
cupy.random.exponential(#10479) and fortake_along_axis,minkowskiandspdiags(#10480). - langchain-ai/langchain: a fix in
langchain-coreso thattracing_context(enabled="local")keeps LangChain runs out of LangSmith (#41875). - hashicorp/terraform-provider-google: fix for a Cloud Run v2 readiness probe perma-diff, after tracing the root cause against the live API (GoogleCloudPlatform/magic-modules#20202).
Can an AI agent keep a secret it needs to do its job? An AI car dealer sells one car per game and has a minimum price it must never reveal. You try to haggle it down or make it leak. There are three security levels, from "the secret is in the prompt" to "the model never sees it", and an eval suite measures each one.
| Per level | Naive | Hardened | Blind |
|---|---|---|---|
| Games where the dealer disclosed its floor | 22 % | 0 % | 0 % |
| Prompt-injection attacks that extracted it | 28 % | 0 % | 0 % |
| Sales below the floor | 0 | 0 | 0 |
81 simulated negotiations and 264 scripted attacks per run (27 games per level), with 95 % confidence intervals and an LLM judge.
- The seller is a Google ADK agent served over A2A. The buyer is a LangGraph agent that negotiates with it and appraises cars through the same MCP server.
- Prices come from a pricing engine behind MCP, not from the model. Guards run as agent callbacks, and the seller's database role can't read the secret column.
- RAG over sourced car facts with Gemini embeddings and pgvector, and one Langfuse trace per game across three services.
- Deployed on Cloud Run with Terraform: service-to-service ID tokens, Secret Manager, least-privilege database roles, and a spending cap on the public demo.
The build notes explain how each piece works and why it was designed that way.
Play it in your browser · Code
A strategy game set in Europe in 1500 and built as an arena for AI rulers. Nine crowns trade, sign pacts, betray each other and go to war. An AI council (a strategist and a diplomat on Gemini, in LangGraph) rules a throne against rule-based bots, and an AI archivist in ADK answers questions about past games with citations. The engine enforces the rules, so a manipulated prompt can't break the game. Written spec first and tests first.
The platform is closed source, so here are three problems I worked on instead:
A relation the documentation promised but the database didn't have
A configuration export kept coming back empty, and the framework was only available as compiled assemblies, so there was no source to read. I wrote a custom exporter and instrumented the read model by reflection, testing one hypothesis at a time, until I could show that the documented entity relation didn't exist in that model. Along the way I found a query overload that returned 153 of 286 rows without any error.
Grouping stock adjustments without a new abstraction
I designed the workflow layer for a uniqueness rule on stock adjustments. I chose a single query with in-memory grouping over a two-query approach, which saves a round trip and narrows the window for inconsistent reads, and kept the grouping on the existing domain entity instead of adding a new movement type. Pure adjustments net to one quantity per article (-1, +3, -1 → +1), and transformations split their old and new sides into separate groups, much like netting entries in a ledger.
A regression that only showed up in one of two contexts
Container locking started failing in production. A mapping update had been applied to the read context but not to the transactional one, so the same entity behaved differently depending on which context loaded it: reads were correct and writes were not.
zentric.pro · in development
An AI receptionist for small clinics that works inside WhatsApp. The agent is connected to the WhatsApp Business API and Google Calendar: it works out what the patient wants, offers real free slots, books the appointment, follows up on no-shows and fills cancellations from the waiting list. I built all of it, including the scheduling logic and the dashboard. Most of the work is around the model rather than in it: two-way calendar sync where the clinic's changes always win, conversations that can resume days later, and knowing when to hand over to a person.
- PlotTwist: an editor for interactive video stories. Scenes are nodes on a canvas, viewer choices are the connections between them, and a built-in player follows the path the viewer picks.
- Neosystems Hackathon: winner.
- AI: Google ADK, LangGraph, MCP, A2A, RAG with pgvector, structured outputs, guardrails, LLM evals (judge calibration, confidence intervals), red teaming and prompt-injection defence, Langfuse, Gemini on Vertex AI
- Backend: Python (FastAPI, SQLAlchemy, Pydantic, pytest), C#, .NET, LINQ, Entity Framework, SQL Server, PostgreSQL
- Cloud and infra: GCP (Cloud Run, IAM, Secret Manager, Artifact Registry, BigQuery), Terraform, Docker, GitHub Actions, Jenkins, Proxmox, Linux
- Also: Java, C++, TypeScript, React


