MLflow + Weights & Biases

What it is

MLflow handles model lifecycle — tracking, registry, deployment. W&B handles experiment dashboards — gorgeous loss curves, hyperparameter sweeps, comparison tables. I use both, depending on what the team already runs.

How Vaaani uses it

• Tracking every training run with hyperparameters and metrics
• Model registry: staging → production with one click
• Hyperparameter sweeps with parallel agents
• Comparing runs across teams visually

Why it makes the cut

Without these, after 50 experiments nobody knows which model is in production or how it was trained. With them, audits and rollbacks are trivial.

Sample code

import mlflow

with mlflow.start_run():
    mlflow.log_param("lr", 0.001)
    mlflow.log_metric("f1", 0.92)
    mlflow.pytorch.log_model(model, "model")