Open for staff / principal roles · Madrid
Retrieval, agent runtimes, guardrails, evals — the infrastructure around the model, built for production teams who need it to hold up after launch.
01 / positioning
I am a production AI engineer with 20 years in software. Before shifting to AI platform work, I spent years shipping code for banks and retailers where production bugs were expensive. My focus isn't prompt engineering or clever tricks to bypass limitations. I build the operational layer around the model: retrieval quality, deterministic runtimes, strict guardrails, and interfaces operators can actually trust when a prediction is wrong.
Full-stack architecture and platform execution behind financial and B2B products.
At Atlax360, I lead AI engineering across BIFROST (our document intelligence layer), ORVIAN (our multi-tenant workflow runtime), and Polaris (our internal knowledge assistant). These are not portfolio mockups or hackathon prototypes—they are live systems running inside financial operations. My job is to ensure they hold up under production stress.
I use coding agents and language models heavily inside my own engineering loop, but I have no illusions about their limitations. The final standard is always deterministic safety, inspectable traces, and software that keeps running when input quality, permissions, or API latency get messy.
Led engineering teams and major rewrite processes inside strict, highly regulated financial environments.
Managed cross-functional platform teams in complex shipping schedules and legacy software transitions.
IBM Machine Learning Professional & Linux Foundation Node.js developer verified certifications.
02 / current proof
Most of my work stays behind enterprise firewalls. These three systems show the technical architecture I build: ingestion pipelines, multi-tenant workflow runtimes, and guardrailed assistant interfaces.
document intelligence
Document intelligence for messy real-world inputs: 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 product combining BIFROST retrieval with guardrails, citations, streaming UX, suggestion revalidation, and operator analytics — one surface for support, sales, and product teams.
03 / open source
Most AI frameworks rely entirely on prompts for behavior control. I publish Rust and TypeScript libraries that enforce determinism, tracing, and PII sanitization at the compile or platform level.
Policy-as-code permission harness for AI coding agents: deterministic rule files instead of prompt-only safety instructions.
Local-first trace recorder and inspector for agentic workflows: append-only evidence, hook ingestion, and signed audits.
Runtime-agnostic structured logging with context propagation, observability primitives, and automatic PII sanitization.
04 / what ships
To me, AI engineering means owning the full execution loop—not wiring up a thin API wrapper and calling it a platform. I build the retrieval index, the orchestration runtime, the evaluation harness, and the operator-facing interface.
Chunking strategy, metadata, permissions, source quality, citations, caching, and the unglamorous work of saying “I don't know” when the index is wrong or the document is ambiguous.
Tool boundaries, stopping conditions, traces, handoffs, evaluators, queues, and cost controls — not infinite autonomy theater that breaks the first time an edge case hits production.
Repository context, executable plans, browser checks, CI guardrails, and review loops around Codex and Claude — so agent output stays legible inside a real codebase.
Next.js, TypeScript, streaming UX, operator screens, and the interface someone uses when the model misfires — because AI products fail in the UI long before they fail in the benchmark.
05 / operating model
I operate with high autonomy and own production outcomes. Based in Madrid, I work remote-first across European time zones, targeting Staff and Principal roles in teams that treat language models as core product infrastructure—not as a marketing layer on top of brittle glue code.
06 / questions
routing guide
Quick facts, core stack, and operational fit compiled for recruiters, hiring managers, and search crawlers.
07
I am looking for teams treating language models as product infrastructure — where retrieval quality, cost, latency, and safety are treated as software engineering problems.