I build AI systems in Python: NLP, speech recognition, LLMs and RAG, from the model to an app people can use.
Advanced Master of AI, KU Leuven.
What I built and what it does.
Problem: the EU AI Act is 144 pages long. Most teams only want to know one thing: is our system high risk or not?
Solution: describe your system to the chatbot and it gives you the likely risk level, with the exact pages of the law behind every answer, so you can check its work. It is a first look, not a lawyer.
Live demo, every claim traceable to its source page
Problem: typing subtitles by hand is slow and nobody's favorite job. Most tools only cover a few languages.
Solution: drop in any audio or video file and get .srt subtitles back, with optional translation to English and a mode for vertical video. It runs on your own computer, so your files stay with you. More than 220 people have downloaded it so far.
220+ downloads, 99 languages, live on Hugging Face Spaces
Problem: Latin hexameter, Virgil's meter, has strict rules for every syllable. People have built Latin verse machines since the 1840s, and big language models still struggle with it.
Solution: so I gave it a try in my master's thesis. I trained open language models on 56,000 lines of Roman poetry, with a rule checker that sends every bad line back. Out came Latin verse about love and science, though Virgil would still have notes on the meter.
Master's thesis at KU Leuven, 27 fine-tuning runs, 56,000-line dataset on Hugging Face
Problem: Flemish towns film their council meetings and post each one as a single video of up to four hours. Almost nobody watches them.
Solution: I turned these recordings into one page per agenda item: a short summary, the vote, who spoke, and a link to the exact second in the video. You can search every meeting, in Dutch or English.
3 municipalities, 14 meetings, 26 hours of speech, 286 agenda items placed on the video timeline
Problem: before the EU writes a new law, it asks the public what they think. Dozens of these calls are open at any time, each with a deadline, and organizations in Brussels keep track of them by hand.
Solution: type an organization's name and get one page: the open calls that concern it, what others already answered, who met the Commissioners about it and what the proposal says. A team of AI agents does the reading, every quote is checked against its source, and a reply is only drafted after a human says go.
81 consultations, 9,181 organization submissions read, 98% of model quotes verified verbatim, 40 organization pages, 120 tests
Problem: almost every AI calorie app sits behind a monthly subscription.
Solution: take a photo of your meal or just type it, and the app estimates the calories and macros, then shows what is left for the day. It syncs with Health Connect and Google Health, so your meals go in and your burned calories come back. It is free, it has no account and no server, and I use it every day myself.
Free and open source, syncs with Health Connect and Google Health, four languages
Like to draw, and curious which famous painter you take after? Point your camera at your sketch, or draw one in the app, and it scores the artists you resemble and explains why. No promises it will say Picasso.
Draw-and-analyze canvas, daily painter challenge, share cards
A pipeline that translates all 43,000+ lines of the game's text into a new language, keeps the terms consistent with a 496-term glossary and checks every line before building the mod.
On Steam Workshop, 150+ players subscribed
Use VS Code like Obsidian. Gemd shows your notes as you write them, with links between notes and Kanban boards. Your files stay exactly as Obsidian saved them, so you can use both apps on the same vault.
Free on the VS Code Marketplace
Specialization in NLP and automatic speech recognition.
Thesis: fine-tuned language models paired with a rule-based checker, to write Latin poetry that follows the strict rules of hexameter.
Courses: Machine Learning and Inductive Inference, Natural Language Processing, Speech Recognition, Speech Science, Language Engineering Applications, Fundamentals of AI, AI Ethics and Regulation.
Digital media, ethics and language: how images and text tell a story, and how to take that story apart.
BA thesis on the nature of consciousness in Daniel Dennett and John Searle.
MA thesis in philosophical anthropology on the human condition, read through Nietzsche and Georges Bataille.
Before Leuven, a first BA at Yeditepe University, Istanbul, with a thesis on Newton's and Kant's concepts of space and time.
I started in philosophy, where I wrote about consciousness, what makes us human, and space and time. Then I moved through digital media into AI. The AI Master at KU Leuven made it technical.
My thesis, in computational poetry, is where the two meet. Poetry asks for creative freedom, Latin hexameter demands strict rules, and I wanted to see whether a machine can do both. So I paired a fine-tuned language model, for the creative side, with rule-based logic that checks the meter, and had it write Latin verse on modern themes.
That background shapes how I build. I care about systems that can explain themselves: RAG answers with real citations, generation checked by validators, pipelines that log what they did. I like shipping the whole thing, from the dataset to the Docker container.
Based in Leuven since 2017. If I'm not on a tennis court or out for a run, I'm probably editing video, which is how I got into speech AI in the first place.
Looking for an AI engineer or data scientist in Belgium? I answer fast.