My Work

Projects & Creations

A showcase of personal and professional projects spanning web development, machine learning, and embedded systems.

Contributing to an agentic security pipeline that autonomously detects, exploits, patches, and verifies software vulnerabilities using LLMs.
Primary owner of end-to-end pipeline runs across 120 Java projects from CWE-Bench-Java, producing the scans, exploits, and patches behind the paper's evaluation.
Improved pipeline reliability to run unattended across the full 120-project benchmark by adding graceful fallback that reuses cached scan results and runs fresh analysis otherwise, eliminating prior startup failures.
Documenting the work as a formal research manuscript intended for submission to the ACM Conference on Computer and Communications Security (CCS), emphasizing system reliability, verification, and auditability.

PythonJavaLangGraphLLM AgentsCodeQLStatic Analysis

Led backend development of a 2D tile-based escape game by architecting game logic and AI systems, coordinating code reviews and integration across a team of developers.
Implemented dynamic goblin AI using a modified A* pathfinding algorithm with randomized patrol logic, enhancing difficulty balance and replayability.
Refactored core classes (MapGenerator, CollisionChecker, Entity) to eliminate God classes and duplicated code, improving long-term maintainability following OOP best practices.
Engineered a unit and integration testing suite with JUnit and Mockito, achieving 96% line coverage and 92% branch coverage to ensure correctness across all gameplay interactions.

JavaMavenJUnitMockito

Won the first ML Hackathon hosted by the SFU Data Science Student Society by building a multi-class fraud detection pipeline for fraud detection.
Designed a validation and cleaning pipeline enforcing consistent data types and value ranges, applying constraint-based reasoning to flag real-world transaction impossibilities.
Engineered three predictive features: balance mismatch flag, balance difference (after minus before), and balance error (amount minus balance difference) to expose accounting inconsistencies for the model.
Trained a class-weighted Random Forest over a logistic regression baseline, tuning number of trees, max depth, and minimum samples against Macro F1 to improve recall on rare minority fraud classes.

PythonScikit-learnPandasNumPy