Blog
Field notes from a working AI lab
Research essays, lab diaries, and shorter posts — an illustrative editorial section demonstrating the proposed blog.
19 posts
The EU AI Act at enforcement: a lab's field guide
The European AI Act has teeth in 2026. Here's the working checklist we run first. We use it before a model touches a hospital, a municipality, or a grid operator.
Small language models are eating the enterprise
The frontier-model arms race dominates Twitter. But the real deployments tell another story. In Portuguese hospitals and municipalities, they are almost entirely 3B–14B models.
Reasoning models vs retrieval: when test-time compute wins
Two families of techniques are often pitted against each other. They are retrieval-augmented generation and reasoning models. The truth is they solve different problems.
Clinical LLMs in Portuguese hospitals: the factuality problem
We deployed a Portuguese-language LLM alongside a radiologist. It taught us that factuality isn't an evaluation metric. It's a product requirement.
Concept drift in production: keeping maintenance models honest
Concept drift is the quiet killer of predictive-maintenance models. It strikes on live industrial data. A worked example of the detection stack that catches it.
Safety audits for multi-agent systems: a practitioner's checklist
Multi-agent systems exhibit emergent failure modes. Single-agent systems cannot. Here is the audit we run before one leaves the lab.
Sim-to-real at the edge: the robot-learning bottleneck no one talks about
The sim-to-real literature is full of clever domain-adaptation tricks. The bottleneck we hit in the LIACC robotics lab is simpler. The real robot is slow.
Five LLM prompts we use weekly at LIACC
Five prompts we copy-paste around the lab. The point isn't that they are magic. It's that they give you something to iterate on.
The research engineer's toolbox, 2026 edition
The tools we actually reach for in 2026. Nothing speculative here. All shipping, all in our daily loops.
What I learned debugging a 70B model at 3am
The story of a bad deploy and what it taught me. Here is the short version. Logs you wrote three months ago save future-you.
Inside a RoboCup week — a photo diary
What a RoboCup week looks like from inside a team. A short, illustrative photo diary of a run. And of the year of work behind it.
If you're new to AI research: start here
Maybe you are an MSc or first-year PhD student at Porto or elsewhere. Here is the short path I'd give you. It covers your first twelve months.
Open-weight models we actually ship in 2026
A short, opinionated rundown of open-weight models. These are the ones we deploy at LIACC in 2026. And what we use each one for.
A day inside the Porto Ring Road digital twin
An illustrative tour of a Porto Ring Road digital twin. This is what it looks like from the inside.
Twelve Portuguese NLP benchmarks you should know
A curated list of twelve Portuguese NLP benchmarks. Do you evaluate on Portuguese? One of these is probably on your shortlist.
AI for the rest of us: nine use cases that ship in 2026
Nine concrete AI use cases I've seen ship for real. All in small Portuguese teams. All over the last twelve months.
How I quit writing slides (and picked up a static site generator)
A short, opinionated essay on a switch I made. I replaced Keynote with a static site generator. It now handles most of my internal presentations.
Thesis supervisors want you to know this
Six unwritten rules of writing a good master's thesis. Supervisors wish they didn't have to repeat them. Yet they do, every year.
When to ignore benchmarks
Public benchmarks rank models on the wrong things. That holds for most deployments. A short guide on when to look away.