I'm a BA in Artificial Intelligence student at CUHK working on efficient multimodal models. I study which capabilities of vision–language models degrade, and how, when the visual tokens they see are compressed. I also write a set of notes on markets.

News

  • Jun 2026Launched this site.
  • 2026Working toward a workshop submission on visual-token compression.

Research

  • Visual-token compression for VLMs — merging, pruning, pooling.
  • Capability-specific degradation under a fixed token budget.
  • Omission vs. hallucination — and what causes which.

More on the research page.

Projects

glassbox ↗

Build modern AI from scratch — LLM internals, RL post-training, and vision, one transparent mechanism per file.

Pythonfrom scratch
frame2frame ↗

Head-pose estimation from video — Euler angles (yaw/pitch/roll) with FIR signal smoothing.

Pythoncomputer vision
ItineralTrace ↗

A multimodal AI travel-planning agent with voice, vision, and memory.

agentsmultimodal

Notes

Notes on markets — theses, position rationale, and post-mortems.

All notes →