Rafael Gallardo

I’m an applied scientist. These days I work on AI agents at Oracle, mostly on the tools, guardrails and workflows that help other engineers build agents they can trust.

I started in research at the Language and Knowledge Engineering Laboratory of BUAP, where I also studied computer science. After that I spent a few years on real-time computer vision at Algotive, before moving to language models and agents.

What I work on

• Agentic engineering. How to build software together with AI agents and still understand what got built: written specs, a person who approves each step, and decisions kept as files you can read later.
• Agent frameworks and guardrails. Making agents something a team can describe in configuration and run safely, even when nobody on the team is an ML specialist.
• Inference and infrastructure. On the side, I build and run the infrastructure that serves open models. This taught me a lot about inference engineering, model optimization and quantization, and deploying and scaling models on GPUs, both in the cloud and on our own hardware.
• Computer vision, earlier on. Fast inference on GPUs, small models for cheap devices, and following objects across many cameras.

What I’m curious about

• How people and agents should share work, and who stays accountable for the result
• Open models that run on hardware you own
• Small, efficient models for places with little compute
• Multimodal models that connect language and vision
• Brain-computer interfaces, which I still read about when I can

Things I’ve shared

Papers, all listed on Google Scholar:

• Approaches to EEG-based Brain-Computer Interfaces: A Survey
• Deep Learning for Fast Identification of Bacterial Strains in Resource Constrained Devices
• Evaluación del modelo neuronal de atención visual en la descripción automática de imágenes en Español
• edge2art: Edges to Artworks Translation with Conditional Generative Adversarial Networks
• Facial Recognition using Convolutional Neural Networks and Supervised Few-Shot Learning

Code and data:

• DIBaS dataset: a mirror of a bacterial image dataset whose original website went down
• Spanish Flickr8k: the Flickr8k captions translated to Spanish, with a training notebook
• Efficient bacterial identification: code for the mobile deep learning paper
• edge2art: sketches to paintings with Pix2Pix
• Coeus-AI: semantic search over scanned documents, built with friends for the JusticIA 2021 hackathon

Elsewhere

GitHub · LinkedIn · Google Scholar · ORCID