← all demos ML training automation

Training-Run Automation

An agent chooses hyperparameters, then the server trains a real image classifier (circles vs squares) end-to-end with a live loss + accuracy curve, evaluates on held-out data, and the agent summarizes. Small but real — it demonstrates the orchestration, not large-scale pretraining.

How it works

  1. Plan: the agent picks learning rate + epochs for the task.
  2. Data: synthetic 12×12 grayscale images (circles vs squares) are generated server-side.
  3. Train: a logistic-regression classifier trains via gradient descent (NumPy) — loss + accuracy stream per epoch.
  4. Evaluate + summarize: held-out accuracy is computed; the agent explains the run.
Agent planshyperparams
Trainlive metrics
Evaluate
Summarize