Bryan Huang | Projects

Neural Network from Scratch (MNIST Digit Identifier)

Python • NumPy

Multilayer Perceptron & Hand-Derived Backpropagation

Engineered and implemented a multilayer perceptron (784-16-16-10 structure, ~13K parameters) from scratch using NumPy with no machine learning libraries.
Derived analytical matrix gradients by hand to train with a custom stochastic gradient descent (SGD) loop, achieving ~94.2% test accuracy.
Designed a modular, object-oriented architecture supporting arbitrary layer dimensions and network depths.
Profiled training bottlenecks between full-batch and stochastic gradient descent, discovering parameter update cadence to be the critical factor for rapid convergence.

Python NumPy Neural Networks Backpropagation SGD Optimization

View on GitHub ↗

2023 – 2025

Delaware Bridge Design Competition • 1st Place State Champions

Designed, simulated, and assembled optimized balsa wood bridges across three years of iterative competition.
Constructed 3D beam member assemblies and performed stress and displacement simulations using SOLIDWORKS.
Calibrated finite element models using empirical scaling factors (2.52–3.02) to match real-world physical load tests.
Iterated through 8 distinct CAD revisions and manufactured a monolithic steam-bent curved arch to eliminate joint weaknesses.

SOLIDWORKS FEA AutoCAD 2D Drafting Empirical Calibration Structural Prototyping