Projects

Shuffle is a minimal, single-purpose music player. Audio files live in a /music folder in Firebase Storage; the app fetches the full list on load, picks a random track when you hit shuffle, generates a signed download URL, and streams it through Howler.js. When a song ends the next one plays automatically.

UI Highlights

• Animated vinyl disc that spins during playback
• Audio visualizer with 7 bars that animate while playing
• Seekable progress bar with elapsed / total time display
• Volume slider with mute toggle
• Smooth transitions via Framer Motion and a dark glassmorphism card

How It Works

Song names are derived from filenames — dashes and underscores are replaced with spaces and the result is title-cased. Progress, volume, and playback state are all managed locally in React. The entire player UI and logic lives in a single MusicPlayer.jsx component.

Inventory Alpha identifies inefficiencies in retail supply chains by applying a Recursive Kalman Filter to sales data. Unlike static models, it separates random noise from actual trends, allowing for reduction of excess inventory without sacrificing sales.

The Kalman Advantage

By treating demand as a latent state in a 1D Random Walk, the system accounts for:

• Measurement Noise (R): Filtering out random daily fluctuations and measurement errors.
• Process Uncertainty (Q): Adapting to how fast underlying demand signals actually shift.
• Dynamic Covariance (P): Driving Z-score based safety stocks directly from real-time signal uncertainty.

PatientAware addresses critical inefficiencies in Canada's medical system by providing physician-facing summaries that digest longitudinal patient data.

The system features a dashboard for continuous symptom logging and structured clinical insights, ensuring care continuity. Thread-based patient data storage allows for persistent context across appointments.

Developed at Queen's AI Club (QMIND), this project implemented an ensemble of Reinforcement Learning agents (PPO, A2C, TD3) to manage portfolio allocation adaptively.

The diversification framework significantly reduced volatility compared to standard benchmarks like SPY buy-and-hold, validated through extensive Sharpe ratio and drawdown analysis.

This game assistant uses Bayesian inference to model and update beliefs about the hidden state of the game. It tracks all player suggestions, responses, and seen cards to continuously calculate the probability distribution of the cards in the murder envelope.

Technical Overview

For each card, the system estimates the probability of it being in the envelope given the history of observations:
P(Card ∈ Envelope | History) ∝ P(Card) × P(Evidence | Card ∈ Envelope)

• Inference Rules: When no one refutes a suggestion, the specific cards are immediately assigned a probability of 1.0 (certainty).
• Bayesian Updates: When a player shows a card they previously hid, the system uses a DecreaseFactor to adjust probabilities based on information theory principles.
• Distribution Normalization: Suspect, weapon, and room probabilities each maintain a normalized sum of 1.0 within their respective categories.

Course project for APSC 200: Engineering Design and Practice. Course Grade: A+ | Role: Primary Technical Contributor

Driven by an interest in mathematical optimization, I developed a decentralized multi-agent system designed to tackle the unpredictable nature of wildfires. Moving away from rigid, centralized control, I implemented Lloyd’s Algorithm (i. e. K-Means Clustering) to allow a fleet of drones to "self-organize" based on real-time environmental data.

I built a custom simulation environment in MATLAB where I experimented with:

• Dynamic Spatial Partitioning: Using Voronoi regions to ensure agents automatically distribute themselves to the most critical "hotspots" without human intervention.
• Sensor Fusion Logic: Designing an "Observation Set" function that allows agents to weigh environmental density, effectively giving the swarm a collective "vision" of the fire's intensity.
• Robustness & Scalability: Engineering the system to be decentralized so that the failure of one drone doesn't compromise the mission, as neighboring agents autonomously re-calculate and compensate for the gap.