Applied Mathematics · Machine Learning · Recommendation Systems
I am a Mathematics undergraduate at The University of Hong Kong, pursuing a second major in Computer Science and a minor in Finance. I build recommendation and retrieval systems, develop numerical methods for inverse problems, and evaluate models through reproducible experiments. I am currently seeking internships in quantitative finance, recommendation systems, data science, or machine learning engineering.
- Semantic IDs, residual quantization, Transformer sequence modeling, and constrained generative retrieval
- Multi-channel candidate retrieval, learning to rank, negative sampling, and temporal evaluation
- Reproducible ML experiments with ablations, baseline comparisons, and multiple pipeline seeds
- Numerical methods for inverse problems, reduced-order modeling, and regularization
Python · PyTorch · Transformers · Residual Quantization · MiniLM
A TIGER-style generative recommender that predicts semantic product IDs from item histories, with controlled experiments on Amazon Beauty.
- Built a residual-quantization tokenizer over frozen metadata embeddings, an encoder–decoder Transformer, and trie-constrained beam search over valid catalog codes.
- Evaluated three seeds under global time cutoffs on a 50,000-user Amazon Reviews 2023 cohort: 6.853% mean Recall@100, versus 7.993% for SASRec, on 2,252 eligible warm test users. Reported ranking and new-item limitations alongside the results.
- Exported a portable CPU inference bundle with CLI/HTTP interfaces, five retrieval/fusion strategies, and full-history seen-item exclusion; audited checkpoint hashes, temporal contexts, and saved rankings.
Repository → · Latest results →
Python · PyTorch · LightGBM · DuckDB
An offline retrieval-and-ranking system built on 31.8 million H&M transaction events.
- Combined nine heuristic, collaborative, learned, text, and image retrieval channels in a quota-aware K=500 candidate interface.
- Trained a LambdaRank model on 100,000 customer-week queries under four rolling temporal cutoffs; an untouched future week reached MAP@12 of 0.03448, a 3.94% relative gap from development.
- Ran bounded-memory inference for 1,371,980 customers across 138 shards and submitted final predictions scoring 0.03117 public / 0.03116 private MAP@12 on Kaggle.
C++17 · Information Retrieval · CMake · LLM APIs
A retrieval-augmented document question-answering system with a deterministic C++ search layer.
- Implemented tokenization, a custom ownership-aware binary-search-tree multimap, and an inverted index.
- Applied AND constraints within expanded term groups and unions across groups before grounded synthesis.
- Injected the model client so retrieval tests remain deterministic and credential-free; live generation reads credentials only from environment variables.
Python · scikit-learn · pandas · Matplotlib
A study of personality, stage-fright, and behavioral-outcome prediction on generated data, with separate model-selection and holdout partitions.
- Kept preprocessing and model selection inside development pipelines before one-time evaluation on an untouched 20% holdout.
- Evaluated 381 model-and-feature-subset configurations with a predeclared parsimony rule.
- Selected a three-feature Gradient Boosting model with 0.9651 holdout ROC-AUC, retaining a compact behavioral feature set.
Python · XGBoost · scikit-learn · pandas
A collaborative regression study comparing tree ensembles, regularized models, and principal-component regression on airfare data.
- The final XGBoost model reached RMSE ₹2,319.47 and R² 0.9896 on a 60,031-row held-out same-period test set.
- Examined model-estimated booking-time effects with partial dependence while keeping predictive associations distinct from causal claims.
- My recorded contributions cover code for Experiments 1 and 2 and the interpretation and discussion for Experiment 2.
Recovering a Hidden Heat Source
HKU Summer Research Fellowship · 2026
Heat spreads over time, blurring the shape of its source. My research asks how we can work backwards from noisy temperature measurements to recover the source's location and shape.
- Built numerical models of heat diffusion and compared full-sized solvers with reduced models using proper orthogonal decomposition (POD).
- Studied how training examples affect the shapes a compact model can recover, comparing reconstruction accuracy and computation time.
- Created a visual demonstration of three models recovering the same letter-shaped source from noisy measurements.
I studied how geometry, engineering constraints, and construction cost affect the rail routes we can build. I combined analytical optimization with Particle Swarm Optimization to search for feasible designs. This work was recognized in the S.-T. Yau High School Science Award.
Languages
Python · C++ · SQL · MATLAB
Machine Learning & Data
PyTorch · LightGBM · XGBoost · scikit-learn · pandas · NumPy · SciPy · DuckDB
Recommendation & Search
Generative Retrieval · Semantic IDs · Residual Quantization · SASRec · Candidate Retrieval · Collaborative Filtering · Two-Tower Models · LightGCN · LambdaRank · Negative Sampling · Temporal Validation · MAP/NDCG/Recall
Engineering & Research
Git · CMake · pytest · Jupyter · LaTeX · Numerical Optimization · Experiment Design
- HKU Summer Research Fellow in numerical inverse problems and reduced-order modeling
- GPA 3.98/4.30; Dean's List, HKU Undergraduate Entrance Scholarship, and Lee Shau Kee Scholarship
- Mathematics coursework in probability, stochastic processes, optimization, numerical analysis, and scientific computing
- Email: pingsenchow@outlook.com
- GitHub: github.com/Stanley-Chow
- Portfolio: stanley-chow.github.io
For internship opportunities or research collaborations, please get in touch.