Barça Innovation Hub
More than a Hack
Our team built IpurdIAr, a football analysis assistant combining tracking data, inferred match events and commentary.
Read the projectAI RESEARCH & ENGINEERING
Adrià Flores Albó.
I graduated in Artificial Intelligence from UPC in June 2026. I'm especially interested in how language models work, from their architectures to the way we train and evaluate them.
I enjoy reading research papers, implementing their ideas, and testing how they work in practice. Recently, I rebuilt a reduced-scale version of Kimi K3 from scratch, pretrained it on 6 billion tokens, and independently evaluated the result. I also worked on document AI during my internship at Medidedalia.
I use this website to share those projects and write about what I learn along the way.
A FEW MILESTONES2022 NOW
BSc in Artificial IntelligenceUPC · 2022 – Jun 2026
AI Research Engineer InternMedidedalia · Oct 2025 – Jun 2026
Independent researchJul 2026 – Present
SELECTED WORK
Projects
A selection of research and engineering projects.
RESEARCH ENGINEERING · PROJECT 01
Kimi K3
from scratch.
Rebuilding a modern language-model architecture, one mechanism at a time.
Kimi Delta Attention, Gated MLA, attention residuals and sparse experts, implemented in a reduced-scale hybrid model.
The challenge was making all the pieces work together.
I connected the mechanisms in a 13-layer PyTorch model, implemented recurrent and chunkwise KDA, and tested their numerical behavior.
The final test was training the complete model.
The 295M-parameter model completed pretraining on 6 billion tokens. A separate next-token evaluation yielded a perplexity of 21.14.
MECHANISTIC INTERPRETABILITY · PROJECT 02
What if a model
could unlearn no?
Can we identify a direction associated with refusal?
Compare internal representations and estimate a direction in the residual stream that is associated with refusing requests.
Then change what the model can express along it.
Project selected weights away from that direction and measure the trade-offs in refusal and general capabilities.
WORK
Experience
My first industry role was an AI research engineering internship at Medidedalia, where I worked on model selection, document anonymization and OCR. Here's what I did there.
01 / EXPERIENCE
Medidedalia
AI Research Engineer Intern
OCT 2025 – JUN 2026 BARCELONA · HYBRID
Work on document extraction and processing for sensitive medical and legal documents.
SELECTED WORK
SchemaRouter
For my bachelor's thesis, I developed a budget-aware router that selects a document-extraction strategy based on schema requirements, quality, cost and latency.
Document anonymization
I built and benchmarked a service that detects sensitive information using rules, fine-tuned NER models and LLM support for ambiguous cases.
OCR and vision-language models
I benchmarked OCR and VLM approaches for medical-legal PDFs and integrated a service that can locate extracted text in the source document.
SchemaRouter● ○ ○
01 / ROUTING
Document
Router
Choosing a model also means balancing quality, cost and response time.
02 / PRIVACY
Clinical document
REDACTED
REDACTED
Finding identifying information before documents are processed further.
03 / DOCUMENT UNDERSTANDING
→
OCRVLMStructured output
Comparing document reading methods and their outputs.
MODEL CHOICES ↗ DOCUMENT SYSTEMS
HACKATHONS & TEAM PROJECTS
Awards
I've taken part in a few hackathons. Here are three challenges my teams won, and what we built for them.
2025–2026 · THREE TEAM WINS
Award details
Datathon FME · Smadex Challenge
Smadex Challenge
We modeled seven-day in-app purchase revenue using LightGBM and neural networks with learned embeddings.
Read the projectBitsxLaMarató · BSC Challenge
BSC Challenge
We benchmarked bioinformatics models for a continuous representation connecting two medulloblastoma patient groups.
Read the projectTECHNICAL ARTICLES
Blog
I write about things I've been working on or trying to understand. Mostly model architectures, training methods and the details that don't always fit into a project README.
01 / TECHNICAL NOTEREAD ARTICLE
ARCHITECTURES
Understanding
Kimi Delta Attention
A closer look at the recurrent update, the gates and the triangular system behind chunkwise computation.
21 min read · Read article ↗02 / TECHNICAL NOTEREAD ARTICLE
OPTIMIZATION
Muon from
first principles
From momentum to matrix updates: why orthogonalization changes the direction of an optimization step.
19 min read · Read article ↗GET IN TOUCH
Contact
Feel free to reach out about research, projects, or opportunities to collaborate.
↑ Back to the beginning