Steering with sparse features
RunningHypothesis — Features recovered by a sparse autoencoder on mid-layer activations can steer refusal behaviour more precisely than prompt instructions, without collapsing answer quality.
I built this page with a simple goal: to show, rather than tell, my skills and the projects I've worked on in AI, Machine Learning and beyond.
The idea comes from a difficulty I've run into many times: explaining what I can do without turning everything into a long list of technologies, skills and keywords. Sometimes words aren't enough; other times they even risk making everything look more complicated than it really is.
That's why I decided to build an interactive portfolio, where you can actually see what I've done, how I did it and, when possible, why I chose to do it that way.
So think of it as a kind of personal lab: a place to collect the projects, experiments, ideas and skills I've developed along the way. It's not a definitive snapshot, but rather a work in progress of what I can do so far.
There is, however, one thing I consider more important than any skill or technology: trying to understand how the underlying mechanisms work. It's an approach that takes time and often means going much deeper than necessary… but once you understand the mechanism, many of the ideas for implementing it come almost on their own.
I hope this page manages to tell you a bit more about who I am, how I think and, above all, what I love to build.
I'm not a frontend developer. I don't enjoy doing frontend and, with all due respect to those who do, I have no intention of becoming one. 😄 Claude and OpenAI helped me solely with connecting and polishing the frontend pieces. The backend, the logic, the algorithms, the ML and the AI are all my own work. In short: they make the interface speak nicely, I take care of everything that happens behind it.
What you see in the background isn't just a visual effect. It's a K-Means algorithm running in real time, which I developed in Python. To choose the optimal number of clusters I use the Silhouette Score. In short, even the background has a job to do.
The centre of gravity. Supervised learning on tabular and sequential data, with more time spent on evaluation design than on architecture choice.
Hypothesis — Features recovered by a sparse autoencoder on mid-layer activations can steer refusal behaviour more precisely than prompt instructions, without collapsing answer quality.
Hypothesis — Binary quantisation of embeddings costs less recall than the memory saving suggests, once a small float rescoring pass is added on top.
Hypothesis — Conditioning retrieval on an entity graph beats flat chunk similarity on multi-hop questions, where the answer is never in one passage.
Hypothesis — Diffusion-generated defect crops improve recall on classes with fewer than 200 real examples. Result — they help until roughly 400 real samples exist, then actively hurt precision.