DocuMage Cheque OCR at a Gulf Bank — 99.1% Field Accuracy on 10,000 Cheques per Day

We deployed DocuMage (Dm), our flagship intelligent document processing platform, in a parallel-run configuration alongside the legacy OCR. For the first four weeks, every cheque that flowed through the existing system was also pushed through DocuMage, with no impact to production. The bank's operations team was given a side-by-side dashboard showing the legacy system's output, DocuMage's output, and the eventual human-confirmed truth.

By the end of week four, DocuMage's field-level accuracy on the bank's actual cheque traffic was 98.4%, versus the legacy system's 87.2%. We then ran a tuning sprint — fine-tuning the courtesy amount and legal amount extraction models on the bank's specific cheque corpus, including the Arabic cursive variations common in the GCC. By the end of the tuning sprint, field-level accuracy was 99.1% and end-to-end straight-through rate (where every field on the cheque passed validation without human touch) was 94%.

We also brought in two supporting accelerators. Signature Verifier (Sv) was deployed to perform on-cheque signature matching against the bank's customer signature card library, flagging signature anomalies for fraud review before payment release. Anomaly Detector (Ad) was wired into the post-OCR flow to flag patterns indicative of fraud — duplicate cheque numbers, sudden amount escalations on a single drawer account, or stylistic anomalies in the legal-amount handwriting.

DocuMage runs on a private cloud tenant inside the bank's existing Azure regional deployment, which the bank's regulator has approved for processing payment data. The platform is deployed as a Kubernetes-native stack on AKS, with autoscaling configured to handle the morning processing peak — cheque volumes typically triple between 09:00 and 11:00 local time.

The OCR pipeline is layered. The first pass uses Microsoft Azure Document Intelligence (Form Recognizer) for base text extraction and field localisation. The second pass uses our own fine-tuned ICR model — a vision transformer fine-tuned on roughly 1.4 million cheque images from comparable banking deployments — to handle the handwritten courtesy and legal amount fields. The third pass uses GPT-4o Vision in a constrained, schema-validated prompt to reconcile the courtesy and legal amounts and resolve any disagreement between the OCR and ICR passes. The output from all three passes is fed into a confidence aggregator that emits a single confidence score per field.

Cheques where every field exceeds a 96% confidence threshold are pushed straight to the clearing engine via the bank's existing ESB. Cheques with at least one field below threshold land in an exception queue, where they are presented to a human reviewer with the original image, the extracted fields, and the model's confidence scores. The reviewer's correction is fed back into a continuous-learning loop that retrains the field-extraction models on a weekly cadence.

Integration with the bank's existing Finacle core happens through the bank's ESB layer using a dedicated cheque-clearing connector we built. The connector enforces the bank's existing pre-clearing validation rules, including drawer-account status checks, stop-payment instruction lookups, and same-day clearing windows.