Projects
- Image classification
Calibrated dog breed classifier
Measured a six-point generalisation gap that the standard benchmark cannot show, because its test images sit in the training set of every pretrained model. The classifier itself names one of 120 dog breeds, refuses the photos it cannot answer for, and reports a confidence that matches how often it is right.
- Stanford test, 8580 images
- 93.11%
- real dog photos the gate accepts
- 98.0%
- p95, deployed VPS
- 123 ms
- Project in development
When a fused Triton kernel is worth writing
A fused kernel is easy to make look fast on its own, because a microbenchmark times the operation and not the model around it. Writes two Triton kernels for the memory-bound tail of a transformer and measures them across batch sizes and sequence lengths, since the activations a fusion saves scale with the tokens processed together while the weights it cannot avoid reading do not. The answer is a crossover rather than a speedup. Coming soon.