Projects
25 projects across 12 categories — filter by topic below.
2026
2024
Modern NLP — From Statistical MT to Multimodal Foundation Models
Four paradigm shifts in one semester: IBM Model 1 → attention-based NMT → transformer parsing → LLM fine-tuning → CLIP multimodal retrieval with pragmatic reasoning. Each technique subsumes and extends the previous.
Neural Radiance Fields (NeRF)
Training an MLP to represent a 3D scene as a continuous function from (x, y, z, θ, φ) to (RGB, density). Volume rendering turns the field back into images; the field itself is the 3D model.
Vision Transformer + Masked Autoencoder
ViT classifier achieving 73.5% on CIFAR-10, then self-supervised MAE pretraining boosts finetuned accuracy to 76.8%. Full implementation of patchify, attention pooling, and mask reconstruction.
Fun With Diffusion Models
Sampling from DeepFloyd IF — CFG, SDEdit, inpainting, visual anagrams. Then training a time- and class-conditioned U-Net from scratch on MNIST to learn the diffusion process end-to-end.
Transformer for News Summarization
Self-attention, multi-head attention, and encoder-decoder architecture implemented from scratch. Trained on CNN/DailyMail achieving 35.1 ROUGE-L, outperforming LSTM baseline by 60%.
RNN Sequence Modeling
Recurrent networks from scratch — forward pass, backpropagation through time, and gradient flow analysis. Vectorized NumPy implementation validated to 5e-5 tolerance.
Auto-Stitching Photo Mosaics
Building a panorama pipeline from scratch — Harris corner detection, Adaptive Non-Maximal Suppression, feature matching, RANSAC for homography estimation, and Laplacian-pyramid blending.
Deep Learning from Scratch
Backpropagation, BatchNorm, Dropout, and CNNs implemented from first principles in NumPy — then PyTorch deployment achieving 74.8% on CIFAR-10.
Face Morphing with Delaunay Triangulation
Smooth warping between two faces via point correspondences, Delaunay triangulation, and affine warps per triangle. Plus: population mean faces and caricature generation by extrapolation.
Filters & Frequencies — Edges, Hybrid Images, and Blending
Working in the frequency domain to extract edges, create hybrid images that change meaning with viewing distance, and blend images seamlessly via Laplacian pyramids. Ends with the famous 'oraple.'
Colorizing 1907 Russian Empire Photographs
Reconstructing color from Sergei Prokudin-Gorskii's glass plate negatives (captured 1907–1915) using image pyramids and normalized cross-correlation alignment.
Pacman AI — A Tour of Classical AI
End-to-end AI agent built across four paradigms: A* pathfinding, Minimax/Expectimax game trees, HMM + particle filter tracking, and Q-learning. Each piece solves one Pacman problem; together they cover the full classical AI curriculum.
Image Geolocation with k-NN & Linear Regression
Computer vision system predicting photo locations from visual features. Combines k-nearest neighbors with regression models, achieving 127km median error on global street-view dataset.
17× Faster 2D Convolution: AVX2 + OpenMP
Hand-optimized SIMD kernels with parallel tiling achieve 17× speedup over naive implementation. Deep dive into vectorization, memory patterns, and performance engineering.
32-bit RISC-V CPU from Logic Gates
A complete 2-stage pipelined RV32I processor built from first principles in Logisim — ALU, register file, control unit, and memory system, all hand-wired from basic gates.
Collaborative Filtering Movie Recommender
SVD-based matrix factorization system with 0.89 RMSE on MovieLens dataset. Handles sparse ratings, cold start problems, and scales to 100k+ users through optimized gradient descent.
Decision Trees & Ensemble Methods
From-scratch implementation of decision trees with pruning, random forests, and AdaBoost. Comprehensive analysis of overfitting, feature selection, and ensemble performance on real datasets.
MNIST Neural Network in Pure RISC-V Assembly
A complete 2-layer MLP for digit classification — matrix multiply, ReLU, argmax, file I/O, and all infrastructure — written entirely in hand-coded RISC-V assembly without any library calls.
Neural Network from Scratch
Pure NumPy implementation achieving 99.6% MNIST accuracy through optimized gradient descent, backpropagation, and regularization techniques.
2023
Entropy Wordle Solver
Information-theoretic greedy solver that picks each guess to maximize expected entropy over the remaining word set — averaging 3.92 guesses across 300+ games.
MCMC Cipher Decoder
Metropolis-Hastings breaking substitution ciphers against bigram frequencies — watch garbled text slowly resolve into English over MCMC iterations.
Build Your Own World
A sophisticated 2D world generator with dynamic lighting, line-of-sight exploration, custom tilesets, and interactive minimaps — infinite deterministic worlds from a single seed.
ngordnet — Semantic Evolution Explorer
Interactive tool fusing WordNet's semantic graph with Google Books NGram data to explore how language evolves — query 'transportation' and watch centuries of linguistic change unfold.
Pocket Planet
A 100×100 world carved from Perlin noise, then colonized by simulated plants that mutate, compete, and converge on the terrain they're fittest for.
No projects match this filter.