Python · PyTorch · VGG19
Notebook-based CNN inspection project that hooks into a pretrained VGG19, captures intermediate feature maps during a forward pass, and renders the most active channels as grayscale activation images. Layer 22, channel 159 lights up most strongly for the sample statue image.
Python · PyTorch
Convolutional neural network built from scratch in PyTorch to recognize handwritten digits. Two convolution-and-pooling stages extract patterns from raw pixels — from simple edges to the loops and crossings that distinguish digits — feeding into a dense classification head trained end-to-end with SGD and cross-entropy loss to 98.81% test accuracy.
Next.js · TypeScript · Supabase · Gemini API
LLM-powered knowledge pipeline that ingests raw study artifacts — PDFs, YouTube videos, URLs — and transforms them into structured learning sessions: summaries, flashcard decks, quizzes, and concept maps. Designed around active recall principles with a UI optimized for information density and flow.
Rust
Terrain synthesis engine generating richly detailed world maps from pure noise. Hand-rolled value noise over dual height and moisture fields, resolved against a 9-biome classification table and rendered to PNG — zero dependencies, full algorithmic control.
C++
Ground-up implementation of the classic arcade title — engine loop, collision system, sprite rendering, and game state all authored from scratch with no external game framework.
