updated sep 28, 2026 · day 2

this week

  • first project: a keyword spotting model under 1M parameters, built from scratch on Google Speech Commands.
  • starting with how audio becomes numbers: waveforms, sample rate, and Mel spectrograms.

open to research and research-engineering roles at frontier labs.

focus

  • the mathematics and foundations behind modern models.
  • papers, and building working versions of what they describe.
  • model training, distributed training, distillation, and evaluation.

next

  • train, evaluate, and shrink the model, then release it on Hugging Face (around day 11).
  • project 2: fine-tune a small pretrained speech model for ASR.

this is a now page. I update it when my focus changes.