Casper started by testing whether an LLM could do the job, and concluded a purpose-built, interpretable model would be more accurate and defensible. They engineered 65+ features from document metadata and trained calibrated one-vs-rest classifiers into a Pass/Fail/Caution architecture, so borderline cases route to review rather than auto-failing. To win reviewer trust, an 'AI Assist' layer explains each decision in plain English. A feedback-driven ETL pipeline captures every human decision and feeds it back into retraining, so the model keeps improving against new fraud patterns. The result reached 99% precision in production at a 1.2% false-positive rate — roughly ten times better than the LLM and baseline tools they benchmarked — and cut the manual escalation workload by about half.