Zealos - Document verification with real human review

Zealos checks documents: IDs, business papers, proofs of address. It extracts the fields, cross-checks them and flags fraud signals. I was one of the engineers on it.

Stage one is a fast extraction model that returns every field with a confidence. Confident fields fill the record. The rest go to a reviewer with the exact region of the document cropped and zoomed, and the correction becomes a training example for the next version.

Stage two combines extraction confidence, cross-checks like face match, address validation and sanctions lists, and behavioral signals into one recommendation: accept, review or reject. Rejects are soft, so the user can upload again. Reviews reach a compliance reviewer who sees the full stack of signals and can override with a reason code the model learns from.

Role: embedded engineer

Per-field confidence with human review below threshold

Cross-checks and fraud signals combined into one recommendation

Human oversight in line with GDPR Article 22