Signature detection for lending packets

A checklist, not a model of the loan

Underwriting already knows which exhibits need a signature: the note, the application, a rider. Encode that as data on your side (signatureRequired on a page). Ask the detector only whether a mark is present. Amazon Textract's Analyze Lending API is a different product. It classifies mortgage packages and is priced on the Textract pricing page (the examples there use $0.07 per supported page). This API does not classify documents. The price comparison for signature-only calls is on the Textract alternative page.

Per-page images

This sample matches the image API that is live today. pages is your own list. The endpoint does not take an array of URLs. One request, one image.

async function missingSignaturePages(pages, apiKey) {
  const missing = [];

  for (const page of pages) {
    const response = await fetch("https://signature-detector.com/api/detect", {
      method: "POST",
      headers: {
        "Content-Type": "application/json",
        "x-api-key": apiKey,
      },
      body: JSON.stringify({ imageUrl: page.imageUrl }),
    });
    const data = await response.json();
    if (!response.ok) {
      throw new Error(data.error || `Page ${page.number} failed (${response.status})`);
    }
    if (page.signatureRequired && data.result.count === 0) {
      missing.push(page.number);
    }
  }

  return missing;
}

Parallelize with Promise.all if you need to. The route does not apply its own rate limit. A 402 means the account is out of credits. Stop the batch there instead of failing one page at a time with the same error. Refunds happen when the detector fails, not when you simply run out of credits.

One PDF, pages in the JSON

With PDF input, you do not loop requests. Send pdfUrl or pdfBase64 and read the top-level pages array. Page numbers in that array are 1-based. creditsUsed is the page count, or 0 when the same file is cached. Compare page numbers to your required-page list. Empty detections on a required page is the same hold as count === 0 in the sample above. The field contract is on PDF signature detection, and a Python version is in Detect signatures in PDFs with Python.

Hand the box to a processor

A count is a coarse filter. The processor still needs to see the mark on the right line. Pass bbox through, or crop it with the method in extracting signature crops. Do not infer the borrower's identity from the box position.

Credits

• $10 per 100,000 credits, which is $0.0001 each. An image that is not served from cache spends 1 credit. A PDF spends 1 credit per page.
• 10 credits are added when the account is created.
• $1 minimum purchase. That is 10,000 credits. There is no subscription.
• Credits do not expire. The balance stays on the account until a detection spends it. A failed detection puts the credit back. Repeating an image or an identical PDF returns cached: true and does not spend a credit. On a PDF that hit is creditsUsed: 0.
• Buy credits from the dashboard or estimate a volume on the pricing page.