Álvaro Salvador Huertas

I design tools to orchestrate AI agents and evaluate the software they generate.

View projects

About

I’m a third-year Computer Engineering student at the University of Granada. I develop Python tools that integrate AI agents into workflows for querying documentation, planning, development and code review.

My projects combine Codex CLI, context versioned in Git, structured outputs and independent evaluation. I use acceptance criteria, automated tests and GitHub Actions to check generated changes and retain evidence before integration under human supervision.

Projects

code-review-orchestrator

v1.2.0 released

An agent’s own verdict is not enough to validate a code change. This workflow adds independent checks and human oversight.

I built a supervised Python workflow for AI-assisted code changes. It combines planning with versioned Git references, developer and reviewer agents in separate sessions, independent tests in rootless Podman, execution limits and inspectable records. Changes require human approval before integration.

  • Python
  • Codex CLI
  • Podman rootless
  • unittest
  • JSON

Evidence & limits

v1.2.0 evidence: 860 ordinary tests passed in the merge CI, with 13 opt-in tests skipped. The PR CI passed all 13 real rootless Podman integrations with no skips. The checks use synthetic agents, without model calls.

Release scope, compatibility and validation: v1.2.0 notes & limits (in Spanish).

These results do not establish general reliability or measure an AI success rate.

Benchmark exercises: illustrative flow, not a recorded run.
Developer agent

Proposes a code change

Independent tests

Evaluate the candidate in rootless Podman

Only with passing tests and valid evidence
Reviewer agent

Inspects the code and evidence

Python controller’s decision

Approve without integrating, request a correction or stop

Up to 2 attempts in total. A correction needs remaining budget and valid evidence; technical errors stop the run.

Logs retained at each stage

Data Crime Scene

Public code and demo

An interactive machine learning lab for investigating data leakage and comparing evaluation procedures.

Guided, Laboratory and Detective modes combine real training on synthetic data, eight case files with evidence and interventions, and reproducible JSON reports. The project puts experimental design, model evaluation and automated testing into practice.

  • Python
  • scikit-learn
  • Streamlit
  • Plotly
Data Crime Scene’s Guided Mode: mode selector and initial case brief, with the prediction objective and timing. Interface in Spanish.
Guided Mode · real application screenshot, before running the experiment. View full-size screenshot (in Spanish)

ShiftLab

Public code · local application

How do image classifiers behave when their inputs change?

I built a robustness lab comparing five classifiers on handwritten digits under noise, blur and contrast changes, with and without training augmentation. It separates training, validation and test, and provides per-image predictions and reproducible JSON reports.

A small benchmark with synthetic image corruptions; it does not establish production robustness or generalisation to unseen writers. The interface explores archived results without retraining.

  • Python
  • scikit-learn
  • Streamlit
  • Plotly
ShiftLab validation screen comparing five classifiers on clean and noisy digit images. Interface in Spanish; these are development results, not the final test.
Real validation screenshot, captured before the final test. Interface in Spanish. View full-size screenshot

I use Codex to help develop these projects. I review the code and check how it behaves.

Capabilities

AI agents, evaluation & developer tooling

Agent Orchestration

Codex CLI integration and coordination of specialized agents in supervised planning, development and review workflows.

LLM & Agent Evaluation

Evaluation cases and criteria, answer and citation review, acceptance tests and failure analysis.

Context Engineering

Document context, Git-versioned references, separation of instructions from data and validation of structured outputs.

Software Quality & CI

Testing with unittest, acceptance and regression checks, pull request reviews and continuous integration with GitHub Actions.

Programming & Data

Python, C++, SQL and relational databases. Linux and Git throughout the development workflow.

Currently learning

AWS

Introductory training.

LanguagesSpanish & EnglishNative proficiency in both

Contact

Let’s talk.

Have an internship opportunity or a question about a project? Send me an email.