How do you mathematically model an Unstoppable Force hitting an Immovable Object? [P]

Synthetic Data Generation

I engineered an LLM to act as a blind labeler across 2,300+ cross-domain matchups. It only saw character names and their native rule sets, never the underlying stats. This forced my XGBoost classifier to derive its own feature weightings from raw, unbiased outcomes.

Catching a Silent Data Leak

My initial accuracy looked suspiciously great. I audited my pipeline and caught a data leak in my train/test split that was mirroring matchups into both sets. I stripped the leak out, expecting the metric to tank. Instead, it went up-hitting 93% on a clean hold-out. The leak had actually been masking a sharper model.

Explainable AI (XAI)

Raw SHAP values mean nothing to an end-user. I engineered a generation layer that feeds the model's SHAP attributions into an LLM alongside strict domain constraints. The pipeline translates its own mathematical feature importance into a plain-English, logically grounded breakdown of how the conflicting rule sets resolved. It doesn't just output a winner; it mathematically justifies how it navigated the nuance without hallucinating.

Full stack, deployed, and live.

• Repo: https://github.com/aidentejada/anime-versus-ml
• Live Endpoint: https://versus.aidentejada.com