About
I’m an ML research engineer based in Spain. I work on language-model pretraining under fixed compute, on agent systems that keep a person in charge of the decisions that matter, and on Kaggle and ARC Prize competitions. I founded Quanta Labs and I teach at Universidad Monteávila.
What I do now
Founder, Quanta Labs (2025 to present)
Quanta builds AI tools for K-12 schools in Latin America: a multi-tenant learning platform, AI-assisted grading in which the teacher keeps the final word, and offline dictation for teachers. I build it end to end: a FastAPI and Celery backend on PostgreSQL with pgvector, a TypeScript monorepo for the web portals, an Expo mobile app, and the AWS infrastructure in Terraform.
Professor, Universidad Monteávila (2025 to present)
Faculty of Economic and Administrative Sciences, Caracas.
Competitions
AWS Trainium Frontier (3rd in Phase 1, now in the final phase), ARC Prize 2026, and Kaggle competitions in medical imaging and cell tracking. I treat competitions as a way to do research against a referee I do not control.
How I work
- State the hypothesis before the measurement. Every run in my ledgers has its prediction written down first, so a result can contradict me.
- Climb a ladder before paying for the full run. CPU check, one core, a smoke run, a short run, then the expensive one. Most bad ideas die on the cheap rungs.
- Know the noise floor. I measure seed spread and pair noise before believing a difference, and I say which reference a number is measured against.
- Let a reviewer try to break it. Claims go through an adversarial review before anything is built on top of them.
- Publish the failures. Every write-up on this site has a section on what did not work, with numbers.
Background
I studied at Universidad Simón Bolívar in Caracas, where I started in electronics and embedded systems. From 2017 I moved into robotics: SLAM, localization, 3D perception and reinforcement learning in ROS and Gazebo. Later work covered cloud data engineering on AWS (lakehouse and warehouse pipelines, Kubernetes deployments), Go backends, and medical-imaging models, much of it through Udacity programs whose projects are on my GitHub.
Tools
| ML | PyTorch, torch-neuronx and NKI on AWS Trainium, torch_xla on TPU, timm, MLX, SGLang |
|---|---|
| Agents | Google ADK, Strands Agents, Bedrock AgentCore, MCP, Gemini, Nemotron, Qwen |
| Languages | Python, TypeScript, C++, Rust, Go |
| Backend and data | FastAPI, Celery, PostgreSQL with pgvector, Redis, Firestore |
| Cloud | AWS (EC2 capacity blocks, ECS, Lambda, CDK), Google Cloud (Cloud Run, Pub/Sub), Terraform, Docker |
| Robotics | ROS, Gazebo, PCL, MoveIt, OpenCV |