Senior AI Engineer · MLOps · Agentic Systems

I build AI systems that
survive contact with production.

Ten years turning ambitious ML — from FDA-grade stroke triage to LLM agents — into products people actually trust. Two Kaggle golds, one CTO tour, currently at HP's AI Center of Excellence.

shipping LLM agents at HP's AI Center of Excellence
Let's build something →
LinkedIn GitHub Kaggle Hugging Face Medium
10+
years shipping
production ML
Kaggle RSNA
gold medals
€2.5M
EIC grant
secured
CE+FDA
regulated
healthcare AI
AHA
peer-reviewed
in Stroke
01

What I'm doing right now

/agents

Agentic systems

Typed pipelines and tool-calling with Pydantic AI, MCP and Google ADK. Structured outputs, agent orchestration — built to be inspected, not just demoed.

/eval

Making AI measurable

Golden sets, regression checks, tracing with MLflow, Langfuse and Logfire. A/B tests with real latency and cost targets — so "better" is a number, not a vibe.

/ship

Prototype → product

Python, Docker, Kubernetes, AWS EKS, CI/CD. The unglamorous production ownership — observability, failure modes, deployment — that keeps AI alive after launch.

02

I'd rather show you than tell you

A mix of the serious and the curious. I communicate best by building — so here's the receipts.

FLAGSHIP · HEALTHCARE AI · CTO
2016—2024

Methinks AI — stroke triage that clinicians actually use

MethinksLVO and MethinksNCCT: deep-learning detection of large-vessel occlusion and intracranial hemorrhage on non-contrast CT, in real acute-stroke workflows. CE marked, FDA cybersecurity certified, multicenter clinical validation published in AHA Stroke. I led the technical org — architecture, delivery, and a team of ten — while staying in the code.

Deep learning CE / FDA Clinical validation €2.5M EIC grant
HP · AGENTIC ASSISTANT

NIO

An internal AI assistant for industrial-print customers. Agent orchestration plus the evaluation pipelines that keep its answers honest in production.

TENYKS · MLOPS

CLIP retrieval, +30%

Fine-tuned a CLIP retrieval model to lift image retrieval by 30%+, on a platform built with Flask, Celery, MongoDB and AWS EKS — measured, not guessed.

PERSONAL · LIT × NLP
read ↗

Rayuela, computationally

A computational reading of Cortázar's novel: multilingual embeddings, stylometry, LLM-perceived style, 5,000 permutation tests. Turning ambiguous language into testable structure — the same habit as the day job.

OFF-GRID AI · PLAY
try it ↗

Small Cuts

A deadpan AI narrator for your life. An 8B vision model watches a first-person moment; a tiny open voice speaks the slightly-too-honest truth. Only small open models, no cloud.

WEARABLE · PLAY
open ↗

Brass Wings

A steampunk micro-flight game for Meta Display Glasses: one-scan install, canvas vector craft, browser-native procedural audio. Proof that I'll chase a weird idea to the metal.

INFRA · HOME LAB

NemoClaw

A multi-machine home lab — DGX Spark (GB10), Mac Studio M4 Max, a Pi mesh — for test-time training, MCP server work and multi-agent RL. So I stay close enough to the metal to reason about it.

03

The path so far

2024 — now

Senior AI Engineer — HP, AI Center of Excellence

LLM systems for internal engineering teams: agentic workflows, evaluation pipelines, production observability. Leading NIO, an AI assistant for industrial-print customers.

2024

MLOps Engineer — Tenyks AI

Production platform infra (Flask, MongoDB, Celery, AWS EKS), LLM/agent components with MCP, and a CLIP retrieval model fine-tuned for a 30%+ lift.

2016 — 2024

Chief Technical Officer — Methinks AI

Built and led the technical org behind CE-marked, clinically-validated stroke-triage AI. Two Kaggle RSNA golds along the way (2019, 2020) and a €2.5M EIC grant.

education

ESADE MBA · Universidad Simón Bolívar, MSc CS (Magna Cum Laude)

A humanistic-technical profile — business judgement and computer science, pointed at the same problem.

↳ the back room

The Multiverse of Carlos is still running.

My original arcade site — dodge bugs, hit PANIC, get eaten by ghosts. Because not everything needs a business case. It's about the vibes.

▶ press start
04

Let's talk.

Remote preferred, open to relocating for the right thing. I care more about what we're building and why than where the desk is — purpose is the part I won't compromise on.

Barcelona · EU work authorization · Spanish & English

✉ macayaven@gmail.com LinkedIn ↗ GitHub ↗