H&M Two-Stage Personalized Recommendation System

Flagship · Individual project

Built an offline retrieval-and-ranking system over 31,788,324 transaction events.

• Unified nine heuristic, collaborative, learned, text, and image retrieval channels behind a quota-aware candidate interface with a target of 500 candidates per query.
• Trained LightGBM LambdaRank on 100,000 customer-week queries from four rolling target weeks, using cutoff-safe features and group-aware negative sampling.
• Reached MAP@12 0.034479 on an untouched future week (3.94% below the development score), plus 0.03117 public / 0.03116 private Kaggle MAP@12.
• Generated predictions for 1,371,980 customers in 138 bounded-memory shards; 1,362,281 received personalized output and 9,699 used a documented fallback.

This is a portfolio-scale offline system, not a production serving stack. The repository documents retrieval ablations, temporal validation, reproducibility controls, and remaining work.

View repository

Behavioral Personality Analytics

Individual project · Classification

Designed a leakage-resistant study of personality and behavioral-outcome prediction.

• Compared 381 model-and-feature-subset configurations using development data only.
• Selected a compact three-feature Gradient Boosting model that achieved 0.9651 ROC-AUC on the untouched holdout set.
• Reached 0.9910 holdout accuracy on the associated stage-fright task.
• Kept exploratory selection separate from final evaluation and documented the limits of generated data, external validity, and causal interpretation.

View repository

Flight Price Prediction & Pricing Analysis

Collaborative project · Regression

Contributed the code for Experiments 1 and 2 and the interpretation of Experiment 2 in a six-question airline-pricing study over 300,153 rows.

• Compared polynomial regression with XGBoost for fare prediction.
• XGBoost reached RMSE ₹2,319.47 and R² 0.9896 on a same-period held-out set of 60,031 rows.
• Analyzed route, class, stop, airline, and timing premiums while distinguishing predictive associations from causal claims.

The split is random rather than temporal, so the reported results should not be read as forward-market performance.

View repository

DocuQuest Agent

Individual project · C++17

Built a retrieval-augmented document question-answering system with a deterministic search layer.

• Implemented tokenization, a custom ownership-aware unbalanced BST multimap, and an inverted index.
• Applied AND constraints within each expanded term group and unions across groups.
• Injected the model client so retrieval tests remain deterministic and credential-free; secrets are read from the environment.
• Used CMake and focused tests to exercise index ownership, lookup, retrieval, and orchestration.

The current system is lexical: it does not include embeddings, a learned reranker, or guarantees against hallucination.

View repository

Lemmings Game Engine

Course project · C++17

Implemented the actor and world logic for a tick-driven 2D game engine within a supplied course framework.

• Designed an actor hierarchy with polymorphic state transitions and object lifecycles.
• Coordinated collision, terrain, spawning, goals, hazards, and skill effects on a 20×20 data-driven grid.
• Kept responsibilities separated between actor behavior and world orchestration.

Framework code and media assets were supplied; my implementation is concentrated in the Actor and StudentWorld components.

View repository

Parabolic Inverse-Source Reconstruction

Ongoing research · HKU Summer Research Fellowship

Developing numerical experiments for reconstructing indicator sources from noisy diffusion observations.

• Distinguishes raw source-space error from heat-visible error at an observation scale and geometric error of thresholded interfaces.
• Studies regularized reconstruction, level-set stability, and how diffusion changes what can be identified from data.
• Builds coverage-aware proper orthogonal decomposition with offline-online diagnostics for snapshot-law shift.

Manuscript packages are in preparation and are not presented here as published work. See the research overview for the current scope.

Urban Rail Network Optimization

Independent research · 2023

Modeled urban rail alignment under geometric, curvature, feasibility, and construction-cost constraints. Combined analytical derivations for simplified cases with Particle Swarm Optimization for multi-intersection routing. The work was recognized in the S.-T. Yau High School Science Award.