Projects
NietzscheGPT
What happens when you fine-tune a language model on the complete philosophical works of a single author? Fine-tuned GPT-2 Medium (355M parameters) on Nietzsche's entire English corpus — 25 books, ~2.2M tokens — to generate philosophical text in his unmistakable voice; trained on Apple M3 Pro via MPS backend in ~4 hours and published on Hugging Face.
Transformer from Scratch
Building understanding from the ground up — implemented the full Transformer encoder from first principles, bridging mathematical derivations and working PyTorch code in a modular notebook.
- Sinusoidal positional encoding, multi-head self-attention, position-wise FFN, residual connections, layer normalization — all from scratch.
- Trained on a synthetic autoregressive task to verify correctness end-to-end.
10D Hypercube Visualization
Geometric structure rarely survives projection. This project constructs a 10-dimensional hypercube and systematically studies what each dimensionality reduction method preserves and destroys.
- 1,024 vertices; 5,120 edges — visualized under PCA, t-SNE, and layer-stratified views.
- Comparative panels illustrating preservation and distortion across projections.
Live Emotion Detection
A real-time 7-class facial emotion classifier designed for high-throughput inference without sacrificing temporal stability.
- DeepFace with MTCNN detection; temporal smoothing via 10-frame sliding window.
- Frame-skipping for throughput, overlaid confidence bars and FPS counter.