City University of Hong Kong
I am an undergraduate researcher in Data Science at the City University of Hong Kong and a research-track Student Ambassador at the NVIDIA AI Technology Center Hong Kong. I study neural architectures at the operator level: what each operator does to its inputs under the data's native measure, rather than where its parameters sit in a coordinate frame. That view covers problems which otherwise look separate, including context conditioning, low-bit quantization, MoE routing, model compression, and behavioural bounds. It replaces weight-space distance proxies with operator action as the invariant a transformation must respect, so a modification carries a guarantee rather than an empirical hope.
Education
-
City University of Hong KongB.S. in Data ScienceAug. 2024 - Jun. 2028 (Expected)
Experience
-
NVIDIA AI Technology Center (NVAITC)Student Ambassador (Research Track)Jun. 2025 - Present
Honors & Awards
-
100% Academic Scholarship2024
-
Dean's List2024
-
Diversity Grant Recipient2024
-
Tiger Programme Member2024
News
Selected Publications (view all )
MaRK: Markov-adapted Recurrent Kernels for Dynamic Operator Conditioning in State Space Models
Syed Ibrahim Omer, Ginny Y. Wong, Xiangyu Zhao
Advances in Neural Information Processing Systems 39 (NeurIPS 2026) — Poster
State Space Models (SSMs) offer an efficient alternative to Transformers for sequence modeling, yet conditioning pre-trained SSMs for iterative generation typically operates outside the recurrent operator, through input injection or activation modulation. While such mechanisms expose the model to conditioning information, they leave the underlying temporal dynamics fixed. We introduce MaRK (Markov-adapted Recurrent Kernels), a dynamic operator-conditioning framework that maps context vectors directly into bounded modulations of a frozen SSM's recurrence ($A$), read-in ($B$), read-out ($C$), skip ($D$), and discretization ($\Delta$) parameters. Viewed through the lens of Linear Parameter-Varying systems, MaRK induces a context-indexed family of Markov parameter sequences, allowing each diffusion timestep to reshape the model's input-output memory kernel. We instantiate MaRK on a frozen 111M-parameter Hydra SSM backbone and study three adapter geometries: Hypernet, Chebyshev polynomial, and Discrete Cosine Transform kernels. Since these adapters modify the Markov parameter sequence through low-rank auxiliary maps on the frozen backbone, parameter-efficient fine-tuning arises as a structural consequence of the adaptation mechanism itself, requiring only 6.3--11M trainable auxiliary parameters to transition from a bidirectional objective to an iterative diffusion regime. The bounded recurrence parameterization further yields an analytic Affine Quadratic Stability certificate for the modulated recurrence. Through synthetic LPV recovery experiments and Markov-operator diagnostics, we show that MaRK recovers coordinate-invariant temporal operators under matched assumptions and produces distinct, stable timestep-conditioned memory profiles. Empirically, the Chebyshev variant yields the strongest performance, achieving an average validation loss of 2.55, followed by the DCT (2.59) and Hypernet (3.77) geometries. Together, these results provide initial evidence that dynamic operator modulation is a principled operator-level conditioning mechanism for adapting SSMs beyond input-stream injection and adaptive normalization.
Redefining Hybrid Blockchains: A Balanced Architecture
ArXiv 2025 Cited by4
Blockchain technology has completely revolutionized the field of decentralized finance with the emergence of a variety of cryptocurrencies and digital assets. However, widespread adoption of this technology by governments and enterprises has been limited by concerns regarding the technology's scalability, governance, and economic sustainability. This paper aims to introduce a novel hybrid blockchain architecture that balances scalability, governance, and decentralization while being economically viable for all parties involved. The new semi-centralized model leverages strategies not prevalent in the field, such as resource and node isolation, containerization, separation of networking and compute layers, use of a Kafka pub-sub network instead of a peer-to-peer network, and stakes-based validator selection to possibly mitigate a variety of issues related to scalability, security, governance, and economic sustainability. Simulations conducted on Kubernetes demonstrate the architecture's ability to achieve over 1000 transactions per second, with consistent performance across scaled deployments, even on a lightweight consumer-grade laptop with resource constraints. The findings highlight the system's scalability, security, and economic viability, offering a robust framework for enterprise and government adoption.
All publications
Projects (view all )
Migration Surge Prediction: An Early Warning System
PythonSDSC 2005 · Introduction to Computational Social Science · City University of Hong Kong
An early warning system for US visa-issuance surges that fuses 170k+ news articles, Google Trends interest and real exchange-rate signals across 15 origin countries into one panel. News is embedded, clustered and labelled on quantized TensorRT engines, and a horizon-aware ensemble forecasts surges one to six months ahead.
NVIDIA Stock Analysis Dashboard
PythonStreamlit and Plotly dashboard with LLM-backed signals
Fuses ten years of NVDA price history with 7,000+ collected news articles into sentiment-driven signals, technical overlays and hypothetical trade analysis. Sentiment runs through FinBERT and VADER, the technical layer through the indicators-cli package, and the narrative layer through an LLM.
SMS Scam Detection: MLOps Pipeline
PythonTracked, reproducible classification experiments
An end-to-end pipeline over 5,574 labelled SMS messages that compares five classifiers, with every run's hyperparameters, metrics and artefacts tracked in MLflow. The best model separates scams at 94% accuracy with a low false-positive rate.
Portfolio Analysis and Optimization
PythonGE2260 · Introduction to Finance · City University of Hong Kong
Constructs a 60/40 fixed-income and growth portfolio from five years of price history, then locates the efficient frontier with Markowitz mean-variance optimization and 20,000 Monte Carlo paths. The growth sleeve prices 18.3% expected annual return at 11.5% volatility, and the blended portfolio is projected to beat the S&P 500 on return and drawdown alike.