No. 01 · Personal
AMSTERDAM · 2026
Independent AI Researcher Amsterdam
Open to research academia or industryWorld models · LLMs · vision · medical AI · on‑device AI
The limits of languagemean the limits of the world? Perhaps, yet we dwellin more than we can name.
§00 · Currently
Built across LLMs, vision, world models, and medical AI. Now I want to go deeper, with passion and faith.
I've shipped research with Tencent, Gradient Networks, MeetaVista, TU Delft, McGill, Tsinghua, and IMDEA. The throughline I care about: useful, durable AI, especially where access and reliability matter.
Day-to-day
An independent AI researcher in Amsterdam. I currently collaborate with industry (Tencent, Gradient Networks, MeetaVista) and academia (TU Delft, McGill, Tsinghua) on cost-efficient large language models, optimization-as-reasoning, and intent-aware world models. Earlier roles include Marie Skłodowska-Curie Fellow, AI Research Engineer at TU Delft Imaging Physics, and an MSc degree from TU Delft.
Two things drive this work: pushing AI beyond text into perception and action, and making those systems fairer and more reachable in practice.
Systems that perceive a scene and imagine what happens next. The bridge from describing the world to acting in it.
Making capable models cheap, dependable, and useful enough to actually deploy. Routing, optimization, and agent design.
Vision as a channel for human signal, not just object detection. Reading emotion, intent, and attention from what people see.
Models that work in the real clinic, not only on the leaderboard. Smaller, faster, and fair across patient populations.
Perception and language models that run locally, not in the cloud. Distillation, quantization, and the craft of keeping small systems capable.
How crowds price information, shift positions, and reveal preferences over time. Reading collective signal from messy human data.
Where AI stops being a model and becomes a tool. Designing systems that fit how people actually think, type, and decide.
A motivation, not a sub-field. Anything I build should reach the people who need it most, not the people who already have everything.
§01 · Research
What runs through every paper here. I want this work to be useful where it matters, accessible to people who don't already have it, and meaningful enough to keep at for years.
ACM CAIS · RLEVAL WORKSHOP · 2026
A step-level benchmark for agentic LLM routing: at each call, pick the cheapest model tier that still resolves the task from the prefix seen so far. A fast static track of 970 execution-verified labels across five workloads pairs with a live dynamic track on SWE-bench Verified, where a trained router matches unrouted Opus 4.6 while cutting API cost by 53%.
ACM CAIS · RLEVAL WORKSHOP · 2026
A trace-driven framework that lets an agentic system evolve three tracks together: a SkillBook of recurring prompt signatures, a learned per-call router, and a small-model adapter, each update admitted only through verifier-backed joint replay. The Skill → LLM → Router schedule reaches 87.3% router accuracy and trims serving cost to 51.8% of always using the large model.
★ IMWUT / UBICOMP · 2025
First large-scale public eye-tracking dataset for VR emotion recognition, with high-frame-rate periocular video plus 240 Hz gaze, across seven discrete emotions.
MICCAI · + IEEE TRANSACTIONS ON MEDICAL IMAGING · 2025
A physics-driven framework that estimates tissue properties from annotated cardiac MRI via diffusion models, enabling zero-shot segmentation across unseen MRI sequences.
MIDL · 2025
Trained nnU-Net models contain substantial weight redundancy. Over 80% of weights can be removed by simple magnitude pruning while preserving segmentation quality.
§02 · Projects
Where the research meets real systems. Some are live collaborations (Gradient Networks, Tencent, MeetaVista); others I build on my own time. Most are still in motion.
Mining Polymarket's order books to study how prediction markets price news in real time. Looking at when crowds get it right, when they get blindsided, and what the order-book dynamics reveal about collective belief.
A framework that uses an optimization harness (structured search, feedback, memory retrieval) to guide LLMs toward better solutions on optimization problems like TSP, rather than relying on direct generation alone.
Building on the LLM Router benchmark, using stronger-model trajectories to fine-tune small specialists for repeated workflows. As usage accumulates, small models improve and the system's cost falls.
A vision-language world model that infers customer intent from visual cues and responds with sales-relevant reasoning, fine-tuned on a synthetic dataset distilled from classical sales literature.
Built a benchmark for evaluating LLM routing strategies. Showed that routing can preserve same-quality performance while reducing cost by >90% versus calling a single SOTA model for every step.
Reframing fine-tuning as generative parameter prediction. A U-Net diffusion model forecasts future weight states from early checkpoints; short, medium, and long-horizon forecasts merge into the final model, yielding 3.2× wall-clock speedup with downstream performance held.
A browser extension that surfaces venue-quality badges, author-role signals, and flexible ranking data directly on Google Scholar. The information you wish was there at a glance.
§03 · Writing
Notes on research, taste, and the boring parts of building AI.
First entry coming soon.
§04 · News
§05 · Teaching
§06 · Contact
I read everything. Reach out about research, collaboration, supervision, or just to say hi. I answer faster than is professional, and I bring my whole self to the conversation.
Open to research academia or industry
Independent AI Researcher
World models · LLMs · vision · medical AI · on‑device AI
signed · T.
