Notes on systems, ML, and research work.

A note on low-resource speech data, speaker variability, and why augmentation choices can matter more than model swaps.

A backend write-up with a before-and-after request-path design, async handoff pattern, and the metrics behind the shift.

A staged NLP pipeline with a gate classifier, NER layer, structured JSON output, and low-confidence review routing.

A CUDA write-up with benchmark crossover points, shared-memory tiling, and the profiler clue that exposed the next bottleneck.