Lucas González Fiz — AI Engineer

Hi, I'm Lucas

I build end-to-end AI systems that hold up in the real world.

AI Engineer at Possible Inc, with a degree in Artificial Intelligence from the Universidade de Vigo. I've built LLM/RAG pipelines for clinical data at MicroPort, real-time perception for a 1:10-scale autonomous race car, and my own agentic, retrieval, and computer-vision systems — with a focus on reproducible evaluation, inference optimization, and deployment.

View projects Download CV ↓Email me →

Selected systems

Projects

Featured project · 2026 · Solo · agentic systems platform

Governed multi-agent enterprise platform

A multi-agent platform for enterprise data with role-scoped tools, read-only SQL access, human approvals that survive restarts, and end-to-end audit traces, evaluated offline on real documents and 541,909 real transactions.

Real transactions: 541,909
Docs faithfulness: 0.94
Approval gate: 1.00

Python
LangGraph
FastAPI
MCP
PostgreSQL
Ollama
Docker
Kubernetes

Case study →Repo ↗

2026
Solo · computer-vision system
Real-time vision for worker-safety monitoring
Industrial fall detection that runs live on a single mid-range GPU — pose, tracking, segmentation and a temporal model behind one observable service, with every figure measured on real hardware.
- Python
- YOLO26-Pose
- SAM 2.1
- ByteTrack
- PyTorch
- TensorRT
- FastAPI
- Docker
Case study →Repo ↗
2026
Solo · LLM / retrieval system
Clearance-aware GraphRAG
GraphRAG that enforces access clearance inside retrieval, so a restricted fact never reaches the model. Dense + BM25 hybrid search, graph expansion, and reranking; the hybrid lifted exact-identifier recall from 0.79 to 1.00.
- Python
- LangGraph
- Qdrant
- Neo4j
- BGE-M3
- Qwen3
- FastAPI
- MCP
Case study →Repo ↗
2026
Bachelor's thesis (TFG) · Universidade de Vigo
Partial observability in deep RL
My bachelor's thesis: a reproducible, multi-seed MiniGrid benchmark of PPO, A2C, DQN, and RecurrentPPO under partial observability. Curriculum learning proved the most reliable lever; explicit memory did not pay off.
- Python
- Gymnasium
- Stable-Baselines3
- PPO
- A2C
- DQN
- RecurrentPPO
- MiniGrid
Case study →

Technical skills

Programming & ML

Python
SQL
PyTorch
scikit-learn
Hugging Face
Stable-Baselines3

GenAI & LLMs

RAG
GraphRAG
LangGraph
MCP
Qdrant
Neo4j
Reranking

MLOps & engineering

MLflow
FastAPI
Docker
Kubernetes
Linux
Git
ROS
NATS
NVIDIA Jetson

Computer vision

YOLO
OpenCV
SAM
ByteTrack
Pose estimation
Tracking
TensorRT

Professional & extracurricular

Experience

Sep 2026 — Present

AI Engineer · Possible Inc · Vigo

Working on AI and automation projects for public- and private-sector clients.

Mar 2026 — Jun 2026

AI Engineer Intern · MicroPort CRM · via Cardiovascular Gallega

Built an LLM/RAG pipeline to normalize heterogeneous clinical tables and map terminology across medical sources, combining retrieval, prompting, and validation.
Developed and evaluated a machine-learning model to detect pacemaker lead failures and support clinical review workflows.

Nov 2024 — Present

Software Advisor · Auria Formula Student AI · Student association

Previously: AI Software Developer · Autonomous SystemsNov 2024 — Jul 2026

Developed real-time perception software for a 1:10-scale autonomous racing platform in an international student engineering team using Scrum.
Trained and compared YOLO cone detectors using pretrained weights and training from scratch; applied structured pruning and prepared TensorRT export for embedded deployment.
Integrated cameras and perception services into reproducible Linux, Docker, Git, ROS, and NATS workflows targeting NVIDIA Jetson.

Background

Education

2022 — Jul 2026

BSc in Artificial Intelligence · 240 ECTS, EQF 6

Universidade de Vigo · ESEI

Foundations in algebra, calculus, statistics, optimization, algorithms, software engineering, databases, networks, and parallel and distributed computing; specialized in machine learning, reasoning under uncertainty, NLP, information retrieval, agents, Big Data, and computer vision.

Bachelor's thesis (12 ECTS): a reproducible, multi-seed benchmark of PPO, A2C, DQN, and RecurrentPPO in MiniGrid, evaluating memory, intrinsic motivation, and curriculum learning. Curriculum learning lifted success on DoorKey-8x8 from 0.37 to 0.95.

Highlights

Milestones

2026Placed 12th of 41 teams at Formula Student UK AI, Silverstone — my second time competing with Auria.
2026Podium in the Inditex Tech Challenge at HackUDC, in a team of four: retrieving the catalog products worn in outfit photos, with YOLO garment detection and a fine-tuned fashion CLIP model.Repo ↗
2025First Formula Student UK AI at Silverstone, working on perception and deployment for the team's 1:10-scale autonomous car.

A bit more

About

I'm an Artificial Intelligence graduate from the Universidade de Vigo, interested in the gap between a model that works in a notebook and a system that works on real hardware.

My work spans LLM/RAG systems and agents, applied machine learning for healthcare, MLOps, and computer vision for autonomous racing and industrial safety. The common thread is end-to-end engineering: reproducible evaluation, optimized inference, clean deployment, and an honest account of where systems fail.

Today I do that at Possible Inc, on AI and automation projects for public and private clients.

Based in: Ourense, ES
Focus: LLMs · CV · MLOps
Languages: Spanish (native); Galician; English (B2)

Get in touch

Contact

Open to collaborations and conversations about AI engineering.

Email is the fastest way to reach me, and I read every message.

lucasglezfiz04@gmail.com Download CV ↓