Home CURRICULUM VITAE / 2026

Research & engineering

Adrià Flores Albó.

AI Research Engineer Barcelona, Catalonia

Research-oriented AI engineer working across LLMs, Computer Vision, and production AI systems, with experience building and pretraining language models from scratch, fine-tuning, OCR/VLMs, document AI and evaluation.

Download PDFPreview CV Get in touch
Scroll to explore
THE PATH SO FAR
01
BSc in Artificial IntelligenceUPC · Barcelona
2022–2026
02
AI Research Engineer InternMedidedalia · Barcelona
2025–2026
03
Independent researchLLM architectures · Pretraining · Evaluation
2026–now
01 / EXPERIENCE

Experience.

Applied research in production-facing AI systems.

PROFESSIONAL EXPERIENCE

AI Research Engineer Intern

MedidedaliaBarcelona · Hybrid

Oct 2025 — Jun 2026

Research and engineering for document understanding and privacy-sensitive AI workflows. Completed my BSc thesis during the internship.

01

SchemaRouter

Developed a budget-aware router for schema-conditioned document extraction, selecting extraction strategies under quality, cost and latency constraints.

02

Document anonymization

Built and benchmarked an anonymization service for clinical and legal documents, combining deterministic rules, fine-tuned NER and LLM support for ambiguous cases.

03

OCR & visual understanding

Evaluated and integrated an OCR service for sensitive medical-legal PDFs, benchmarking OCR/VLM approaches and enabling spatial text localization for source citations.

02 / EDUCATION

Education.

2022 — Jun 2026

BSc in Artificial Intelligence

Universitat Politècnica de Catalunya (UPC) · Barcelona

Academic performance: 8.94/10 (Year 3) · 9.65/10 (final semester)

SELECTED COURSEWORK

Large Language Models (Honors), Advanced Reinforcement Learning, Unsupervised Learning, Neural Networks & Deep Learning, Computer Vision, Voice & Dialogue, Advanced Databases

03 / SELECTED WORK

Selected work.

Projects where I could investigate a mechanism, build something, and measure the result.

01 / PROJECT↗

Kimi K3 from Scratch

Independent Research · Jul–Oct 2026

Reimplemented the Kimi K3 language architecture from scratch as a 295M-parameter hybrid model, integrating KDA, Gated MLA, Attention Residuals, and Stable LatentMoE. Conducted controlled learning-rate and batch-size experiments across three model scales, optimized GPU training and checkpoint recovery, and completed pretraining on 6B FineWeb-Edu tokens.

LLM PretrainingScaling ExperimentsHybrid AttentionSparse MoEGPU Optimization
02 / PROJECT↗

Uncensoring LLMs with Abliteration

UPC · Research project · May 2026

Implemented a training-free steering technique that identifies a refusal direction in the residual stream and removes it through weight orthogonalization. Evaluated refusal behavior and general capabilities on Llama 3 8B.

PyTorchTransformersInterpretability
03 / PROJECT↗

Aspect-Based Sentiment Analysis with LLMs

UPC · Research project · 2025–2026

Implemented a bilingual aspect-based sentiment analysis framework comparing four strategies for small language models: zero-shot prompting, few-shot prompting, synthetic-data augmentation, and LoRA fine-tuning. Built structured prediction and evaluation pipelines for English and Spanish restaurant reviews.

PEFT / LoRALLM Fine-tuningFew-shot LearningData AugmentationNLP Evaluation
04 / PROJECT↗

Graph Neural Networks for Retrieval

UPC × Nuclia · Feb–May 2025

Improved retrieval accuracy from 6.2% to 42.3% by replacing inefficient knowledge-graph traversal with a graph neural network approach; compared three retrieval models.

GNNsKnowledge graphsRetrieval
05 / PROJECT↗

Explainable AI for Clinical Decision Support

UPC · Mar–May 2025

Architected a Seq2Seq autoencoder for 30-day hospital-readmission prediction on a 100k-patient diabetes dataset, and built explainability workflows combining SHAP, feature engineering, and RAG-assisted clinical explanations.

Seq2SeqSHAPFeature EngineeringClinical AI
04 / SKILLS

Tools & methods.

A working toolkit rather than a proficiency scorecard.

01

Machine learning

PyTorchLLM pretrainingScaling experimentsHybrid attentionSparse MoETransformersHugging FacePEFT / LoRAComputer visionGNNsReinforcement learningWorld modelsGenerative modelsDiffusion models
02

LLMs & vision systems

LLM evaluationModel steeringMechanistic interpretabilityQuantizationRAGNEROCR / VLMsvLLMOllamaStructured extraction
03

Engineering

PythonCUDAGPU optimizationC / C++SQLFastAPIDockerKubernetesApache SparkVector & graph databases
05 / AWARDS

Awards.

Three team wins, across different problems and domains.

2026
Winner · Barça Innovation Hub

More than a Hack

Team-built football analysis assistant integrating tracking data, inferred events and live commentary audio.

3,500+ applicants
2025
Winner · Barcelona Supercomputing Center

BitsxLaMarató · BSC Challenge

Benchmarked bioinformatics models for a continuous representation connecting two medulloblastoma patient groups.

06 / LANGUAGES

Languages.

CatalanNative
SpanishNative
EnglishFluent · C1
LET'S TALK

Get in touch.

I'm happy to discuss research ideas, technical work or opportunities to collaborate.

Email me GitHub Download CV