An independent AI researcher in Amsterdam. I work with industry partners (Tencent, Gradient Networks, MeetaVista) and with academic groups at TU Delft, McGill, and Tsinghua. Three things keep me at the desk. Cost efficient large language models. Optimization that becomes reasoning. World models that read intent.

Before Amsterdam I was a Marie Skłodowska-Curie Fellow at IMDEA Networks in Madrid. Before that, a Research Engineer at TU Delft Imaging Physics. My MSc is from TU Delft, in Computer and Embedded Systems Engineering.

Mostly I want to build AI that is useful, that holds up, and that reaches the people who need it.

_{Open to collaborations, PhD opportunities, and research roles. Either side, academia or industry.}

A multifaceted eye tracking dataset for emotion recognition in VR. Periocular video at 120 fps. Gaze at 240 Hz. 26 participants across Ekman's seven basic emotions.

_{The first dataset with high frame rate periocular videos. Four times the gaze frequency of prior work. Open Unity collection and Label Studio annotation tools, all included.}

_{Tongyun Yang†, Bishwas Regmi†, Lingyu Du, Andreas Bulling, Xucong Zhang, Guohao Lan}

Over 80% of weights in trained nnU-Net models can be removed via simple magnitude based pruning. Proxy Dice stays above 0.95 across multiple medical segmentation tasks.

_{Validated on four medical datasets, both 2D and 3D. Critical weights cluster near the encoder and decoder ends. Bottlenecks turn out to be heavily prunable.}

_{Tongyun Yang, Yidong Zhao, Qian Tao}

🜨  World models
_{Systems that perceive a scene and imagine what happens next. The bridge from describing the world to acting in it.}

◉  Large language models
_{Making capable models cheap, dependable, and useful enough to actually deploy. Routing, optimization, agent design.}

◐  Computer vision
_{Vision as a channel for human signal, not just object detection. Reading emotion, intent, attention from what people see.}

♥  AI for medicine
_{Models that work in the real clinic, not only on the leaderboard. Smaller, faster, and fair across patient populations.}

⛨  AI safety
_{Building things that behave reliably when they leave the lab. Robustness, evaluation, the unglamorous work of trust.}

⌂  Fairness × access
_{A motivation, not a separate field. Anything I build should reach the people who need it most, not the people who already have everything.}

♁   _{Originally from Sichuan. Now in Amsterdam, by the canals. Cooks Sichuan in a wok at home. Dried chilis, garlic, peppercorns. Never a covered pot.}

☉   _{Plays tennis with a Babolat Pure Drive (his wife uses a Wilson Blade). Half marathon PB 1 h 43 m 53 s. Aiming for sub 1 h 40 m.}

✿   _{Reads Sartre and the Boom Latinoamericano (Borges). Buys tulips at the Bloemenmarkt every spring.}

✎   _{Planning to adopt 瓜子, a 狸花猫 (Chinese mackerel tabby). Writes longhand with a fountain pen before any keyboard gets involved.}

Email is the fastest way through. I read every one. For research context, the personal site has more.

The door is open for collaborations, PhD opportunities, and research roles. Either side, academia or industry.