Jean Pierrevan Driel.
Developer building apps where AI is the foundation,
not the feature. Looking for a startup to grow with.
How much has a model actually read about this?
DataSubjects is a study I'm running on estimating the per-subject composition of an LLM's training data — how often a given subject was sampled — from the outside, with calibrated uncertainty rather than vibes.
The primary estimator needs no access to the model at all. A BPE tokenizer's ordered merge list is a fossil record of the corpus it was built on: each merge constrains the mixture of data the tokenizer saw, and solving those constraints recovers absolute byte-share proportions from nothing but the shipped tokenizer.json — zero inference calls, zero cost. A second, behavioural estimator calibrates signals that scale log-linearly with exposure (cloze accuracy, Min-K% logprobs, perturbation deltas) against models whose training data is public and countable.
The two estimators measure different things — what went into the pool versus what the model was actually trained on — so their disagreement is the interesting signal. It reverse-engineers a data-curation changelog: which subjects got upsampled, and which got filtered out.
Written as a design document of record first — estimands, access regimes, and an experiment matrix where every experiment has a kill criterion agreed in advance. Two have passed so far; the next one runs on Pythia checkpoints.
Nothing here measures a closed model's corpus directly — that can't be done. The method is proved on models where the answer is checkable, and the accuracy lost as access is removed is reported as the error bar.
I grew up in Peru and moved to the Netherlands to study law — until I started building GPT pipelines to help write my essays, before ChatGPT existed. That curiosity changed everything. I did a data science bootcamp at Le Wagon, then launched my first startup: AICOS, an AI customer service tool for hotels in Peru that signed its first paying clients. Back in the Netherlands I took on freelance web work, won a hackathon building an ADHD reading extension (€500, first place), and enrolled in ICT at Fontys University. Five years of building — first out of curiosity, then professionally, now as a student who keeps shipping on the side.
AI interests me because it removes barriers. Tasks that used to require hours of searching can now be handled — and automated — in plain English. I don't think of AI as a feature you bolt onto an app; I think of it as the layer that makes the whole thing work differently. That said, I build just as comfortably without it: my foundations are in UI/UX design and data science — cleaning pipelines, preprocessing, training ML and DL models, the full arc from messy data to polished interface.
Outside of code I play tennis, train jiu jitsu, hike when I can, and think a lot about wine — specifically the growing and making of it. Jiu jitsu especially makes sense to me as a developer: it's problem solving under physical stress, where the solution has to be found in real time with incomplete information.
My bigger goal is to build my own startup. Right now I'm looking for an internship at a small team — somewhere I can contribute to everything, learn fast, and understand what it actually takes to ship something people use. If that sounds like your company, I'd love to talk.
Building something
and need an extra hand?
I'm actively looking for an internship — ideally at a small startup where I can contribute across the stack and learn what it really takes to build a product. Open to other opportunities too.