$ whoami
A meta-model that watches a sentiment transformer and predicts when it's about to be wrong — not just what it predicts, but whether to trust it.
How the pipeline actually works, end to end. Three independent models vote on sentiment; a fourth model watches for disagreement and predicts when the main model (DistilBERT) is likely to be wrong.
DistilBERT is strong, but not infallible — wrong roughly 1 in 20 times here. Rather than improving the base model, this system predicts when it's about to fail, using signals it can't see itself: cross-model disagreement, confidence calibration, and known hard patterns like negation or intensifier words used as praise.
Built from the SST-2 sentiment dataset with 45% adversarial noise injection (typos, casing, punctuation), blended with hand-written negation and sarcasm-adjacent examples targeting patterns the base sampling under-represented.
Found a bug, have a question, or just want to say the gauge animation is satisfying? This goes straight to my inbox.