Back to the workAnimesh Kumar

04   Work  ·  Adobe

Model optimization framework

Role
Member of Technical Staff
Dates
2021 — 2022
Outcome
−43.5% inference time, −52% cost

Built the optimization path for models moving from research into Adobe products — quantization, batching and serving changes measured against a fixed quality bar. Inference time fell 43.5% and cost fell 52%, which is what made several downstream features viable at all.

Built with

  • Python
  • PyTorch
  • Model serving
  • AWS