lenatriestounderstand
An independent ML research-and-writing site: long-form notes with runnable experiments and interactive visualizations, from LLM post-training and mechanistic interpretability to GPU internals and causal inference.
Personal projects
torchisotonic
Batched isotonic regression for PyTorch tensors, fast on the GPU. Fits every row of a tensor at once, on CPU or with CUDA kernels, with the same results as scikit-learn: calibrating classifiers and monotone fits inside models, from a few elements per row to a hundred million.
pip install torchisotonic 
sharada
A small encoder that makes typed decisions about text in one forward pass: the options come with the request, and what comes back is a calibrated probability for each of them. Nothing is generated or parsed, and the model answers questions it has never seen.
pip install sharada 
torchrolling
Pandas-style rolling and exponentially weighted statistics for PyTorch tensors, fast on GPU. O(1) work per element for any window, fused Triton kernels on CUDA, autograd and torch.compile support, and the same numbers as pandas.
pip install torchrolling 