MLOps · Machine Learning Engineering · LLM Agents · Medellín, Colombia 🇨🇴
Production ML and AI gen in financial services · hyperspectral remote sensing research on the side
I build artificial intelligence and machine learning systems that reach production and survive there.
I work on business problems across the full model lifecycle: feature pipelines on BigQuery, training and batch inference on Vertex AI, orchestration with Airflow on GKE, and CI/CD under a SAFe + DevSecOps setup.
The part I enjoy most is the unglamorous layer underneath — reproducibility, cost control, observability, and the internal tooling that lets a team ship models without heroics.
| Production | Migrating online endpoints to BatchPredictionJob, event-driven retraining, champion/challenger patterns, and pipeline observability |
| Tooling | MLOps scaffolding and internal AI tooling — reusable project templates, automated CI/CD wiring, and a curated catalog of reusable capabilities |
| Agents | Designing and building LLM agents: tool use, MCP integrations, retrieval-backed workflows, and the evaluation harnesses that keep them honest |
| Research | Detecting phosphorus stress in Phaseolus vulgaris from UAV hyperspectral imagery — multiclass classification over Zarr datacubes |
| Learning | Rebuilding the LLM timeline from scratch: 2017 Transformer → scaling → RLHF → RAG → agents, one runnable stage at a time |
Languages & Core
Cloud & Orchestration
ML & Data
GenAI & Agents
MLOps & DevSecOps
| Project | What it is |
|---|---|
| ResearchOS | A build-to-learn lab for AI systems: ingests arXiv, PubMed and RSS, answers questions over the corpus with RAG, and ships a personalized morning briefing. Hand-written against provider SDKs — no frameworks, minimal AI assistance — because the goal is to own the fundamentals, not to ship fast |
| spectralcrop-thesis-master | Master's thesis — phosphorus deficiency detection in common bean from UAV hyperspectral imagery, with a fully reproducible ML/DL workflow |
| gcp-learning-vault | Study notes in Obsidian (Feynman technique) on the Google Skills tracks, with the goal of earning the Google Cloud Professional ML Engineer certification |
| clean-agents-template | A production-ready Cookiecutter template for building AI Agents and LLM applications using Python and Clean Architecture |
More in the repositories tab.
Open to conversations about MLOps, AI, remote sensing, and anything that makes AI/ML systems less fragile.