BatchBouncer

BatchBouncer is a web app (with optional CLI) that watches model training in real time and automatically detects data and pipeline failures that waste GPU hours: corrupted batches, label leakage, exploding gradients, silent NaNs, skewed class sampling, and augmentation bugs. It plugs into PyTorch/Lightning and Hugging Face Trainer via a lightweight callback, streams metrics and batch fingerprints, and triggers “safe actions” like pausing, lowering LR, switching to a known-good dataloader seed, or alerting a Slack channel with the exact offending samples. The product is not a full MLOps platform; it’s a focused guardrail for teams that already have training infrastructure but keep losing days to flaky runs. Expect some false positives early; the value is in catching the obvious, expensive failures fast and giving reproducible evidence to fix the root cause.