[home]   [research projects]   [research interests]   [publications]

Coding Projects

These are not repository mirrors. Each page is written as a small technical manual: what the program is trying to do, the architecture of the system, the mathematical mechanism, the source code where that mechanism actually appears, and how the experiments are run.

Numerical Computing

• Scalable Interpolation
  Distributed implementations of classical interpolation methods together with a finite-recurrence reconstruction method and a common speed/accuracy/extrapolation benchmark harness.

Machine Learning / Sequence Models

• Transformer Architecture Ablation
  A controlled forecasting codebase in which the same model family can be run encoder-only, decoder-only, or as a full encoder-decoder.
• Complex Memory
  An active sequence-model memory project. The public page gives the problem, background, and toolset; current architecture and experimental details are intentionally withheld.
• Spectral Attention
  Experiments that place spectral structure directly inside attention and in a streaming damped-oscillatory token mixer.
• Synthetic Data
  A local-first corpus construction pipeline that plans from seed examples, builds a curriculum, generates through several job modes, filters aggressively, and then evaluates the resulting corpus.

Applications / Systems

• ShaTranZ
  An Android speech-translation system connecting Kivy, chunked microphone recording, a FastAPI/Whisper backend, multilingual text rendering, and spoken output.

Machine Learning Experiments

• Linear Regression with Mini-Batch Gradient Descent
  From-scratch NumPy gradient descent with initialization, learning-rate, batch-size, and weight-decay experiments.
• Multi-Layer Perceptrons
  Synthetic nonlinear classification/regression with activation, depth/width, optimizer, and regularization comparisons.
• Convolutional Neural Networks
  CIFAR-10 experiments on CNN depth, a ResNet-18 comparison, and SGD versus Adam.
• Recurrent Neural Networks
  Tiny Shakespeare character modeling with RNN/LSTM comparisons, context-length tests, and hidden-size ablations.

The diagrams on the individual pages are part of the documentation, not decorative thumbnails.