I study computer science with a machine learning specialization and spend most of my time at the intersection of systems programming and physics-informed modelling. I care about correctness, low-level control, and models that can be reasoned about — not just benchmarked.
I am a computer science student entering a machine learning specialization at Ontario Tech University. My primary interest is physics-informed neural networks — models structured around physical priors rather than treating equations as external constraints.
My work runs the full stack: from low-level C++17 infrastructure with formal correctness proofs in Lean, to custom ODE solvers and training pipelines in Python. I am drawn to problems where the abstraction boundary between mathematics and implementation actually matters.
Beyond systems and research, I maintain the D++ Discord bot infrastructure at 143,891 peak CCU, built SavePack — a privacy-first Firefox extension for tab management — and led development of Mecha, a Revolt platform bot that reached 20,000+ users.
Outside of code, I write and research local history. I value precision and directness — in prose as much as in systems design.
I want to see a world where generative AI is regulated — where deployment accountability, data provenance, and consent aren't optional features. Building ML systems means caring about what they do when they leave your hands.
skills & tools
A sophisticated dashboard and invite portal for community management. Features a robust OAuth logic layer running on Cloudflare Workers with ultra-responsive Redis caching for global session state. Built for community scaling — depth of integration hidden behind a premium, minimal interface.
A research toolkit for training and evaluating physics-informed neural networks. Includes a standalone ODE solver with analytic and RK4 numeric paths, automatic equation classification, and a CLI layer with structured JSON logging. Designed for iterative research workflows where reproducibility is non-negotiable.
A research inquiry examining the structural gap between physics simulations and real-world behaviour, proposing Physics-Informed Machine Learning as the mechanism to close it. Built a spring-mass-damper apparatus with an eddy-current brake, trained a PIML network with a composite loss enforcing the equation of motion, and quantified improvement over classical simulation. Connects directly to Heisenberg uncertainty, Faraday induction, and the limits of what any model can know.
A privacy-first Firefox extension for tab management. Lets you snapshot any set of tabs as named packs, close them, and reopen everything in a new window or Firefox container exactly when you need it. Supports tagging, starring, group collapse, instant search, and JSON export/import. Nothing stored remotely — everything lives in Firefox's built-in local storage. No accounts, no servers, nothing leaves your browser.
A substantial Discord bot in C++17 using the D++ library, backed by PostgreSQL. Integrates Tesseract OCR for image-to-text extraction and Whisper for audio transcription. Includes a Markov-chain market simulation with daily price distributions, a modular permission system, and Lean correctness proofs for core probabilistic guarantees.
A real-time Keras training visualization dashboard using Matplotlib and background threading. Supports EMA smoothing of loss curves, checkpoint history carry-forward across interrupted runs, and softargmax-based peak estimation. Built to handle the messiness of long training runs without losing state.
Led development of one of Revolt's most widely-adopted early bots. Built with Voltage.py on a nascent websocket-based API with unique scaling constraints. Reached 20,000+ users and ranked #3 on the platform. Archived after the platform's open-source ecosystem matured — the codebase remains public.
Open to internships, research roles, and interesting collaborations. I don't need a pitch -- direct is better.