AI Student @ UPV/EHU · AI Agents · LLMs & MCP · Machine Learning
I'm a 4th-year Artificial Intelligence student at the University of the Basque Country (UPV/EHU).
I recently completed a 3-month internship building production AI agents: multi-agent systems with LangGraph, a computer-use desktop companion in Rust/Tauri, MCP integrations and systematic LLM evaluation.
I'm interested in building and understanding intelligent systems, especially combining different areas of AI such as deep learning, language models and optimization algorithms.
My main areas of interest are:
- 🤖 AI Agents — multi-agent systems (LangGraph), MCP, computer use, tool design and evaluation
- 🧠 Machine Learning & Deep Learning — model training, evaluation and comparison
- 💬 Natural Language Processing — transformers, fine-tuning vs prompting
- 👁️ Computer Vision — pose estimation and neural architectures
- ⚙️ Optimization & Metaheuristics — heuristic search, genetic algorithms
Languages
AI Agents & Backend
ML / DL Frameworks
Data & Notebooks
Tools
Worked on a production SaaS platform of AI agents (private codebase), across the full agent stack: reasoning, tools, memory and evaluation.
- Multi-agent systems with LangGraph: routing, delegation, human-in-the-loop teams, 140+ automated evaluation cases
- Computer-use desktop companion (Rust/Tauri): agents control Windows/Linux apps via accessibility APIs (AT-SPI, UIA/COM)
- MCP integrations, secure credentials in agent skills, natural language → workflow compiler
- Neo4j knowledge graph as agent memory; LLM observability with Phoenix
AI Agents LangGraph MCP Rust Tauri FastAPI Vue.js
Do LLM agents work in Basque? The same 30-task battery run in English, Spanish and Basque — measuring the agentic layer, not just knowledge.
- Basque keeps its reasoning (100%) but loses the agentic layer: tool use drops to 25% without tools, and answers leak into other languages
- New metric: language leaks — plus localized output contracts (ANSWER/RESPUESTA/ERANTZUNA) as a measured variable
- Public leaderboard with failure gallery, n=3 runs with std, code-verified battery
AI Agents LLM Evaluation Basque Low-resource languages Tool use
Measured when agentic machinery actually pays off: the same 15-task battery solved three ways, on two models.
- Direct prompting vs single tool-using agent vs multi-agent (LangGraph)
- Without tools both models fail only SHA-256 — with tools, the $1/MTok model matches the $5/MTok one (15/15)
- Native tool use, MCP server, automatic evaluation: accuracy, token cost and latency
AI Agents LangGraph MCP LLMs Evaluation
Built and compared different architectures for human pose estimation using the MPII dataset.
- CNN (ResNet) vs Transformer (TokenPose)
- Achieved 82% vs 48% accuracy
- Full pipeline: preprocessing, training, evaluation and visualization
- Analysis of spatial precision vs global context
Computer Vision Deep Learning Transformers PyTorch
Explored whether large language models can replace supervised training in classification tasks.
- DistilBERT (fine-tuning) vs LLMs (Qwen, LLaMA)
- ~97% accuracy vs <30% with prompting
- Analysis of bias and limitations of generative models
- Data augmentation experiments
NLP LLMs Transformers Prompt Engineering
Solved a graph optimization problem using different heuristic and metaheuristic approaches.
- Implemented GRASP, Simulated Annealing and Genetic Algorithm
- Genetic Algorithm achieved Q ≈ 0.89
- Compared solution quality, time and stability
- Trade-off analysis: exploration vs exploitation
Optimization Metaheuristics Genetic Algorithm Graph Theory
- Designing and evaluating agent systems in production
- Rust for systems-level and desktop development
- Deeper understanding of optimization techniques
I'm always open to collaborations, part-time roles and interesting conversations about AI.
email: gorkafernandez47@gmail.com