Production AI · full-time, contract, fractional · Madrid · remote EU
Madrid-based AI engineer. Twenty years in software, ML since 2015. I build retrieval, agent runtimes, guardrails, evals, and operator screens for systems exposed to production traffic, permission checks, slow APIs, bad context, and support escalation.
01 / background
I have worked in software for twenty years and in ML since 2015. Before AI, I shipped for banks and retailers, where defects hit money movement, customer data, compliance review, and release windows. That is the lens I bring to retrieval quality, runtime determinism, guardrails, audit trails, and operator interfaces.
Full-stack architecture and platform work behind financial and B2B products, before moving into production AI.
At Atlax360 I lead AI engineering across three systems: BIFROST (document intelligence), ORVIAN (a multi-tenant workflow runtime), and Polaris (an internal knowledge assistant). They run inside live financial operations; I set the retrieval, evaluation, and guardrail standards the three share, and keeping them reliable under load is the job.
I use coding agents and language models heavily in my own work, and I design around malformed input, missing permissions, rate limits, timeouts, and tool failures. The system should fail closed and keep the prompt, retrieved sources, tool calls, permissions, latency, and decision path inspectable.
Led engineering teams and full rewrites inside regulated financial environments and high-traffic retail.
Ran cross-functional platform teams through complex shipping schedules and legacy transitions.
IBM Machine Learning Professional and Linux Foundation Node.js developer certifications.
02 / current work
Most of what I build sits behind enterprise firewalls. These three show the kind of architecture I work on: ingestion pipelines, a multi-tenant runtime, and a guardrailed assistant.
document intelligence
Document intelligence for the documents finance actually runs on: ingestion quality gates, semantic and visual chunking, pgvector/HNSW search, caching, source-quality scoring, analytics, and explicit no-answer paths when evidence is weak.
workflow runtime
Multi-tenant AI workflow runtime: context assembly, durable memory, deterministic and cached execution tiers, queue processing, idempotency, run events, and human-review metadata when automation should stop.
internal assistant
Internal assistant that combines BIFROST retrieval with guardrails, citations, streaming UX, suggestion revalidation, and operator analytics. One surface for support, sales, and product teams.
03 / open source
The patterns behind my production work, extracted into Rust and TypeScript libraries: permission boundaries, trace evidence, governed memory, and structured observability.
Deterministic policy engine for AI coding agents: maps tool calls to capabilities, evaluates YAML rules, and signs every decision in a verifiable audit log, with hard-stops that policy can't bypass.
Self-inspecting, auditable memory for AI agents: surfaces the evidence, provenance, and health behind each recall so callers can see which memory to trust, with an optional Ed25519-signed tamper-evident ledger. Local-first, Rust.
Local-first trace recorder for AI agent runs: append-only, verifiable evidence of what the agent called, what it was allowed, and what failed, with hook ingestion for Codex/OMX harnesses.
Runtime-agnostic structured logging with automatic PII sanitization (GDPR/HIPAA/PCI-DSS) and native W3C tracing. Zero dependencies; runs on Node, Bun, Deno, Edge, and the browser.
Research harness exploring whether a coding-agent harness can measurably improve itself through typed, layered modifications validated against operator-defined evals within strict budgets.
04 / what i work on
I take one production AI path from retrieval index to operator workflow. The scope includes runtime state, evals, guardrails, latency, permissions, and failure handling in the product UI.
Chunking, metadata, permissions, source quality, citations, and caching, with explicit refusal when retrieved sources are weak, contradictory, missing, or out of scope.
Tool boundaries, stopping conditions, traces, handoffs, evaluators, queues, and cost control, defined before an agent reaches production.
Eval sets, abstain logic, and regression traces. The discipline of proving a change helped instead of assuming it did.
Permissions, audit trails, PII handling, and human-in-the-loop review for AI running inside regulated finance, where “mostly right” isn't good enough.
Repository context, executable plans, browser checks, CI gates, and review logs around coding agents, so their output stays verifiable inside a real codebase.
Next.js, TypeScript, streaming UX, and operator screens for model failures, showing retrieved evidence, failure state, audit trail, escalation path, and the next valid action.
05 / how i work
I work async and stay accountable for retrieval quality, runtime behavior, latency, cost, permissions, and operator workflows. Madrid-based, remote-first across the EU. Open to Staff, Principal, Architect, and Forward Deployed roles, and to contract or fractional engagements, where retrieval, agents, evals, and guardrails are part of the shipped product.
06 / questions
Quick facts on stack, roles, and fit. The CV has the full timeline.
Download CV07
I work with teams shipping LLM features that need owned retrieval quality, cost ceilings, latency targets, permission checks, and failure review. Available full-time, contract, or fractional. Madrid-based, remote across the EU.