I build AI systems that survive an audit, not just a demo.

Javier Pontón González is an AI Engineer and LLM Engineer based in Asturias, Spain, working fully remotely for clients in the EU, UK and US. He designs and operates production LLM systems: AI agents and multi-agent workflows with LangChain, LangGraph and MCP, Retrieval-Augmented Generation with embeddings and cross-encoder reranking, structured document extraction with vision-language models, and self-hosted or air-gapped inference. He has 7+ years of production backend engineering behind that, for Iberia, KPN, Mercadona, Inditex, Openbank and Idealista.

Yes. He works as an independent B2B contractor invoiced from Spain, is available immediately, and works fully remotely across EU, UK and US time zones. He takes both long-running engagements and shorter scoped work such as a two-week AI feasibility assessment. He can be reached at javierpontongonzalez@gmail.com or +34 623 920 307.

He built a production Model Context Protocol server and a RAG service over engineering documentation at Iberia, using embeddings, hybrid search and cross-encoder reranking, instrumented with OpenTelemetry and Dynatrace. He architected a two-agent system for Arabic public-tender response for a Saudi systems integrator, combining structured extraction from vendor quotations with RAG and Arabic OCR over RFPs of hundreds of pages. He also ships RAG in his own products: Facturias retrieves over Spanish tax regulation, and Grabia retrieves over a local meeting archive.

Yes, this is a core specialisation. He designed and delivered a fully air-gapped multi-agent platform for a Saudi systems integrator under PDPL and NCA requirements: model-agnostic architecture against OpenAI-compatible APIs, benchmarking of Qwen3, DeepSeek, Falcon-H1 Arabic and ALLaM, GPU and VRAM sizing for vLLM and Ollama serving, and no data leaving the client network. His own products run self-hosted vision-language models for the same reason, including GDPR special-category data in ZORRO.

Evaluation is defined before the build. A golden dataset representing real inputs, field-level precision and recall as the scoring rubric, regression runs on every prompt or model change, LLM-as-a-judge and human evaluation where the output is subjective, and hit rate and MRR for the retrieval layer. Everything is instrumented with OpenTelemetry so latency, token cost and failure modes are visible per agent step.

Four live products as founder and sole engineer. Facturias (facturias.es) is a multimodal invoice-processing SaaS with a self-hosted vision-language model and RAG over Spanish tax regulation. ZORRO (somoszorro.com) is a dating app for the gay and queer community on iOS and Android with semantic matchmaking and self-hosted photo moderation. Apunta (apuntapp.com) is a multi-tenant SaaS for shooting clubs owned end to end from custom NFC hardware to the mobile app. Grabia is a fully local meeting-intelligence pipeline running on a private inference node.

Python, Java, Kotlin and Elixir on the language side. FastAPI and Spring Boot for services. LangChain, LangGraph and MCP for agents. PostgreSQL with pgvector, Elasticsearch, Redis, DynamoDB and MongoDB for data. Kafka for events. vLLM, Ollama and mistral.rs for self-hosted inference. AWS, Azure and GCP with Docker and Kubernetes. Hexagonal architecture, DDD and TDD as the default way of building.

He is based in Asturias, Spain, in the Europe/Madrid time zone, and works fully remotely with clients across the EU, UK and US. He is a native Spanish speaker with professional working proficiency in English.