Projects

A 2D physics engine that runs entirely inside a fragment shader. Shaders keep no state between frames, so each pixel of a feedback buffer stores one particle's position and velocity, letting the simulation carry state from frame to frame.

• Penalty-method collisions against walls and other particles
• Gravity and linear damping, integrated with semi-implicit Euler
• Click to spawn new particles in real time

Also on Shadertoy.

Predicted how users would rate books from a large interaction dataset, scored on a class-wide leaderboard.

• SVD collaborative filtering, tuned by grid search over latent factors, learning rate, regularization, and epochs
• Separate read-prediction model: a Random Forest trained on balanced positive/negative samples with popularity features

Detected and localized brain tumors in MRI scans, moving step by step from a linear baseline to deep detection models.

• Classification: logistic regression (81.9%) → fully connected network → ResNet-10 with batch normalization (94.0%)
• Localization: fine-tuned Faster R-CNN (ResNet-50 + FPN) to draw bounding boxes around tumors (red = predicted, blue = ground truth)

Designed an instruction set and processor under tight constraints: 9-bit instructions, 8-bit registers, and only 512 instruction slots.

• 14-opcode ISA, a SystemVerilog CPU, and a Python assembler
• Ran 16-bit Hamming-distance and 32-bit multiplication programs on the 8-bit datapath