A fully automated weekly newsletter pipeline built on Google Cloud Platform. Six specialized agents orchestrated via Pub/Sub and Firestore fetch, score, and summarize the top ArXiv AI papers and industry news, then compose and deliver personalized HTML emails to subscribers — with bilingual support, subscriber preferences, and Firebase Auth.
A Reinforcement Learning (RL) agent designed to optimize water release strategies for hydroelectric dams during extreme flood events. The project focuses on autonomous decision-making to ensure water levels remain within critical safety boundaries to prevent structural failure. Please note that the source code is proprietary and cannot be shared publicly, though the full technical report is provided.
A RuneLite plugin that exploits a timing mechanic to grant unlimited prayer points. Reached 40+ paying customers through organic marketing via YouTube videos and forum posts.
Proved that forcing a Reinforcement Learning (RL) AI agent to face a dedicated adversary trained to exploit its weaknesses cuts convergence to a provably unbeatable, game-theoretically optimal strategy nearly in half — from 17.5M to 9.6M training steps — implemented from scratch in a custom multi-agent tactical shooter environment.
Built a benchmark revealing that frontier VLMs — including GPT-5.4 and Gemini 2.5 Pro — systematically fail at imagining 3D perspective shifts, with lateral accuracy collapsing below random chance at large viewpoint misalignments.
Built a deterministic clinical guideline compiler that transforms clinical guideline PDFs into machine-executable, audit-ready decision workflows—using LLMs only at build time. The system enforces strict schemas, line-level traceability, and fail-closed validation, producing reproducible artifacts with cryptographic hashes and zero inference-time AI. Structured reasoning, engineered for safety and trust.
A Sudoku solver that solves any solvable n2 × n2 puzzle, where n ∈ ℤ+, using tree recursion. The algorithm recursively explores possible board configurations, backtracking when conflicts arise. This ensures all possibilities are considered, guaranteeing a solution for valid puzzles. By combining recursion and backtracking, the solver efficiently handles larger grids and adapts to different Sudoku variations.
An application that uses REST API requests to retrieve NHL shot data and convert it into feature vectors for machine learning. It features neural networks, decision trees, logistic regression, extensive feature engineering, and data preprocessing. A dockerized tool allows real-time analysis by inputting a game ID. The codebase includes tools for performance evaluation, graphing, and goal probability estimation.