AI Fraud Document Review

MII Fraud Detection: OCR and two-pass LLM procurement review, with offline development-set evaluation. The hybrid engine remains in shadow mode only.

90.9% to 36.4%
False-positive rate on noisy clean packets, offline development set
76.2% to 0%
Fail-open rate, offline development set
0 of 3,780
Production runs with user-visible output changed by the shadow engine

Role: Creator and sole engineer, responsible for design, AI coding agent direction, review and release
Period: May 2026 to Oct 2026
Project type: Key project
Room in the 3D HQ: L3:fraud-review-platform

Situation

Procurement teams needed timely, trustworthy fraud review across document packets. Early versions flagged clean packets, undermining confidence in the results. As the creator and sole engineer of MII Fraud Detection, I had to show that false positives could fall without missing real fraud.

Role and key decisions

Split review into per-document findings and a cross-document verdict
Why: The pipeline runs over OCR output: the first pass records findings for each document, the second pass reaches a verdict across the documents, and live progress streams while it runs.
Gated changes on gold sets for false positives, recall and fail-open
Why: I first set up gold sets and CI gates for false-positive rate, recall and fail-open, then replaced the majority-vote guardrail and the overly broad safe patterns.
Kept the hybrid risk engine in shadow mode after it failed recall
Why: It cut offline false positives to 9.1%, but recall dropped to 0.841 and it failed its gate. I kept it in shadow mode, where it changed user-visible output in 0 of 3,780 production runs.

Results

90.9% to 36.4%
False-positive rate on noisy clean packets, offline development set
Development gold set.
76.2% to 0%
Fail-open rate, offline development set
Same development gold set.
0 of 3,780
Production runs with user-visible output changed by the shadow engine
Offline false positives fell to 9.1%, but recall dropped to 0.841. The hybrid engine failed its gate and remained in shadow mode only.
11.1% to 3.7%
Prompt-injection success rate in a live evaluation, not production
Live evaluation after a model change; false positives also fell from 26.1% to 21.7%. These are evaluation results, not production outcomes.

Lesson

Write the release gates before running the evaluation; if a change fails a gate, it does not become the decision maker, even when another number looks better. One good development run is encouraging, but it is not the bar for letting the engine decide.