FIG. 001 PORTRAIT "I care about representations that generalize and projects worth polishing."
I'm a third-year undergraduate at The Hong Kong Polytechnic University, supervised by Prof. Bing Wang. I am interested in what representations is needed to achieve visual intelligence. To date, I have explored this question through spatial reasoning via explicit graphs decomposition, and world models where representation naturally emerges. I hope to develop models that can learn predictive representations that allows AI systems to understand and reason about the physical world.
When it comes to projects, I treat each and every of them as a work of art, a few of which I am proud of; a full list can be found in my github repo. Outside of university, I enjoy filming and video production. I once grew a channel past 100K subscribers. I have also independently produced a documentary and an AI short film, with the latter winning an award at the 15th Beijing International Film Festival (BJIFF).
I'm applying to PhD programs starting Fall 2027. I look forward to exploring questions surrounding spatial understanding in greater depth.
P-2026-05 FIG. 014 A general framework for 3D layout generation that improves the reliability and consistency of multi-hop relative spatial reasoning via invariant spatial decomposition, self-consistent imagination, and reparameterized differentiable optimization.
§3.07 FIG. 014 Proposes a unified framework that moves clinical ML beyond binary prediction. Instead of treating classification, clustering, and pattern mining as isolated tasks, I connect them through principled information flow: calibrated risk estimates guide manifold analysis, which in turn generates clinical hypothesis. The framework surfaced numerous clinically validated phenomena.
§3.04 Desktop E2EE messenger implementing full Signal cryptographic stack (e.g., X3DH and Double Ratchet) totalling ~20,000 lines of code. Security properties are formally verified using TLA+ and Alloy, with in-transit ciphertext validated against the NIST 800-22 statistical test suite. This as a course project that far exceeded baseline requirements.
§3.05 Explores the accuracy-latency Pareto frontier in AIoT systems. Existing dispatchers sacrifice one for the other. I implement Lyapunov drift-plus-penalty with auto-tuned V and a time-quantized dynamic programming planner that jointly optimize over the decision space. The DP planner pushes the Pareto frontier furthest, achieving near full-model accuracy with under 1% deadline misses across all tested arrival patterns.
§3.07 Empirically investigated whether pretrained foundation models obsolete both classical methods and manual feature engineering on tabular data. Through comprehensive ablations, our findings: they do, and the best preprocessing is none at all. We also managed achieved Top-10 result on Kaggle's House Price Prediction challenge.