AMY on film.

A real run, recorded. Three of the six sources never get a request: a portal whose terms prohibit automated collection is refused before the fetcher is reached, a host whose robots.txt returns a 500 is closed for the run rather than assumed open, and a host that disallows this crawler by name is obeyed even though it allows everyone else. The rest are swept under their own robots.txt at their own crawl-delay, sitemap first, with the run's filters applied to roster cards so twelve profiles are never fetched at all. Then every licence is checked against the state — which catches one that expired while the brokerage page still advertises the sales — three pages resolve into one Ray Ortega, two Whitfields in one city stay two people, and one pair nobody could settle becomes a question rather than a merge. The last forty seconds are the review console itself, driven from the keyboard. The numbers are whatever that run produced.

The short cut of the same live run: why a referral book is the loan officer's whole business, the query that describes one, the three sources this refuses to read and why, the sweep, the licence the state says expired, and the three pages that turn out to be one person.

Two lanes at the front, because leads arrive both ways: one closed on the phone, one typed on the website. They converge on the same record. AMY reads thirteen documents, chases the one missing page, and submits — and then underwriting pushes back, asking for a signed letter of explanation covering a 21% drop in self-employment income and an $18,400 deposit. AMY relays that to the loan officer and to the borrower at the same time. She replies in two lower-case texted sentences; AMY turns them into a properly structured A4 letter of explanation, files it against the condition, and the loan clears. At the end she can reissue her own pre-approval at any price — and AMY refuses to write one above the limit she actually qualifies for.

The screencast the 1.0 one should have been. A borrower applies, sends thirteen documents including two phone photos, and gets a pre-approval figure derived from what those documents actually say. Then it keeps going into the half that was never on camera: the record it becomes in the AMY CRM, the income and asset work AMY does to it, the write into AMY OpenLOS with the HTTP requests visible, and the package stacked in the order an underwriter reads it. Acts three onward run the shipping pipeline and engines over real bytes — the $498,000 and the $10,474.17/mo are computed live, not written into the page.

A screen recording of the /demo1 page actually running. A loan officer types the application, it becomes a record in the AMY CRM, and fifteen real documents are uploaded — eleven PDFs plus a phone photo, a skewed low-resolution scan, a blurry statement and a fax. AMY parses the PDFs, runs OCR on the images in the browser, catches a shift differential the OCR misread as $1,385.00 and re-derives the true $385.00 from figures it can corroborate, notices four of the uploads are re-sends of documents it already has, then computes income, reconciles the assets, runs the condition rules and files the loan into AMY OpenLOS. Nothing is pre-rendered; the numbers are whatever the engines returned.

The brand reel. Fifteen documents land and get classified in place. A declining bonus is caught and the lower figure governs. Business funds are excluded, a large deposit is sourced automatically. AMY closes eight of nine conditions on its own and answers the underwriter in 21 minutes. The one condition left — a signature only a person can give — reaches the workbench as the single, named exception.