The Evidence Log
Drawn from live case files — what a human or a model would have missed
Real report screenshots from production credit evaluations. Faces are blurred and every name, phone number, GSTIN/UDYAM number, address and internal file path has been redacted before publication — the annotations in red are the AI's own, from the original file.
It read a half-visible sign in a shop photo, then went hunting for that name in the bank statement.
A second, informally-used trade name appeared cropped in the corner of a storefront photo. AI Agent pulled the fragment, searched months of statement narrations for matching credits, and reconciled income a name-only check would have excluded.
It noticed the "shop photos" were screenshots of a screenshot.
Status-bar artifacts and social-media chrome inside the submitted images gave it away: the "live" shop-front, interior and selfie were all screen-recordings of an old video, not fresh captures. Flagged before a human reviewer opened the file.
It read a municipal plaque written in Assamese to confirm the shop and the GPS pin were the same place.
OCR on a background signboard, in a regional script, cross-checked against the declared GST address and the photo's geo-tag — closing a verification gap no standard KYC flow reaches.
It caught a gold-loan disbursal hiding behind a vendor-sounding narration.
A ₹36,000 credit narrated to look like ordinary business income was, in fact, a two-tranche gold loan disbursal once matched against the bureau tradeline — reclassified out of income, with the full ₹51,698 adjustment disclosed line by line.
It figured out what the shop actually sells, and threw out the wrong income model.
Signage, shelf contents and a selfie pointed to a kirana counter, not the AEPS/money-transfer outlet implied by ₹69.7L of pass-through banking volume. Re-basing the model against the real business dropped adjusted income to its true figure — 0.65% of raw credits, not 78%.
It searched the applicant's own phone number and found his DJ business on YouTube.
A stated profession that's hard to verify through paperwork — confirmed in under a minute by matching a declared phone number to a public channel with over nine thousand subscribers and two hundred uploaded event videos.
It read the file path inside a screenshotted photo — and the PC's username matched the business name.
Buried in a photo of an editing workstation: a visible Windows file path and an open project file. The username in that path was the applicant's own declared business name — a corroboration no bureau pull or human glance would have caught.
