Selected projects

Technical work that supports my research ambitions.

These projects reflect how I approach technical work outside formal research: building systems focused on evaluation, interpretability, and real-world usability.

American Sign Language (ASL) Recognition

A real-time computer vision and machine learning system built to explore data collection and robust inference in real-world conditions.

What I Learned

Dataset construction under constraints: built a custom dataset of 2,000+ labeled ASL samples, highlighting real data collection difficulty.

Evaluation beyond static accuracy: validated a TensorFlow model (90%+ accuracy) under varying lighting, pose, and motion conditions, focusing on the gap between test-set metrics and live inference.

Noise-aware inference design: implemented motion-stabilization gating for stable gesture recognition, reducing spurious classifications.

Human-centric ML: designed real-time confidence threshold controls and asynchronous text-to-speech output, focusing on real-world use-cases.

Neural Network from Scratch (Foundational Modeling)

A ground-up implementation of a feedforward neural network designed to make model behavior explicit, inspectable, and explainable.

What I Learned

Mechanistic understanding of learning: implemented forward propagation, backpropagation, and gradient descent directly in NumPy, making parameter updates explicit.

Evaluation without abstractions: trained on the MNIST dataset (95%+ accuracy) using custom architecture design, weight initialization, and manual hyperparameter tuning.

Model analysis: visualized loss curves, weight evolution, and accuracy trajectories across epochs using Matplotlib to observe learning dynamics.

Explanation-first design: structured the project as a code-first, math-supported walkthrough, emphasizing clarity over concision and interpretability over convenience.

Algorithm Learning Platform (Pedagogical Visualization)

An interactive educational platform designed to make algorithmic processes inspectable, understandable, and reproducible.

What I Learned

Interpretability as a design problem: visualized 20+ algorithms and data structures step-by-step, emphasizing clarity and system behavior.

Communication-first controls: designed pause, reset, and inspection tools to support different learning styles and paces.

Correctness-first engineering: built entirely in Elm using functional programming principles, enforcing immutability and explicit state transitions.

Proven instructional use: adopted by UTA faculty and used by 50+ students per semester in live lectures and self-paced study.

Live Demo

Queen's Logic Puzzle Generator & Game

A full-stack puzzle system focused on constraint satisfaction, generation, and verification.

What I Learned

Constraint satisfaction & verification: designed a bespoke flood-fill–based board generator that guarantees exactly one valid solution per puzzle.

System modularity: implemented a FastAPI + Pydantic backend for puzzle generation and validation, decoupled from the SVG-based client.

Operational robustness: deployed as a Docker container with rate limiting, CI/CD, and automated redeployments on GitHub push.

Live Demo