For digital lending

Document processing at onboarding speed

A lending app onboards in minutes, which means the document work happens while the customer is watching. It has to be fast, it has to be right, and the unmasked original must never become a stored object.

Talk to us about your document volume Read the docs

Talk to our technical team → — for integration questions, a security review or a volume estimate.

Mumbaiap-south-1, data stays in India
In memory onlynever written to disk
PDF, JPG, PNGup to 50 MB per document
Batchone call, many documents
Peak-sizedpriced on your monthly volume

Three things that break at app scale

Photographs, not scans. Your users shoot the card on a phone, at an angle, in bad light, sometimes upside down. Detection runs several rotations and several OCR passes precisely because a single clean pass fails on real uploads.
Traffic is spiky, not average. A campaign or a salary-day surge is not your mean load. Per-minute throughput is set per plan and sized to your peak, not to your average, because the average is the number that lets you down.
The retry storm. A worker pool that treats every 4xx as retryable will hammer an endpoint for an hour over a 403 that was never going to succeed. Our errors are typed so your client can tell "back off" from "stop and alert someone".

Where the unmasked copy usually survives

The upload bucket

Files land in object storage before processing and nobody ever deletes them. Mask before the write, not after.

The CDN or image cache

A document served back to the app for preview gets cached at an edge you do not control.

Error logs

A failed parse logs the request body. That body is a document.

The analytics pipeline

Events carrying extracted fields flow to a warehouse that has an entirely different access model from your core database.

Mask at ingestion and all four stop mattering, because the sensitive version never exists long enough to be copied anywhere.

What a phone camera actually sends you

Every assumption that holds for a flatbed scan fails on a user standing in a shop with one hand on the card.

Rotation

Cards arrive sideways and upside down constantly, because the phone was held in landscape. Detection runs several orientations rather than assuming one.

Glare

Laminated cards under a tube light lose an entire digit group to reflection. This is the single most common cause of a zero-count response.

Crop

Users photograph the card at an angle and clip a corner. If the clipped corner held part of the number, there is nothing to find.

Compression

Chat apps recompress images hard. A card forwarded through WhatsApp before it reaches you is materially harder to read than the original.

Design the retake, not just the API call. A zero-count response is your cue to ask the user for another photo while they are still in the flow. Discovering it in a nightly reconciliation means calling someone who onboarded last week.

Sizing for a spike

Quote your peak minute, not your monthly total. Ten thousand documents a month is nothing; ten thousand in the two hours after a campaign email is a different provisioning question, and the monthly number hides it completely.
Name your bursts. Salary day, campaign sends, festival lending, quarter-end. These are predictable and we would rather provision for them than have you discover the ceiling.
Bound worker concurrency to the plan's per-minute rate. A pool sized to your CPU count will exceed the limit and spend its time backing off.
Jitter the back-off. A fleet that retries on the same schedule reconverges into the same spike it just caused.
Watch X-Quota-Remaining and alert early. It is on every response. A campaign that eats a month of quota in a day is visible hours before it becomes an incident.

Questions we get asked

How fast is a single document?

Processing is OCR-bound, so it depends on page count and image quality rather than on a number we would like to print. Send a representative sample and we will measure your actual document mix rather than quote someone else's average.

Can throughput handle a campaign spike?

Per-minute rate is set per plan rather than per tier. Tell us your peak, not your average, and we provision for the peak.

What should our client do when it gets a 429?

Check the error field. rate_limited clears within the minute — back off exponentially with jitter. quota_exceeded will not clear until the period rolls, so alert instead of retrying.

Why did a document come back with nothing masked?

Almost always glare on a laminated card, a clipped corner, or heavy recompression from a chat app. The document is returned unchanged with a count of zero. Treat it as a cue to ask for a retake while the user is still in the flow.

How should we size a plan for a lending app?

By your peak minute rather than your monthly total. Ten thousand documents a month is trivial; ten thousand in two hours after a campaign is a provisioning decision, and the monthly figure hides it entirely.

Talk to us about your document volume

Tell us how many documents a month you process and what your peak looks like. We will size a plan, provision a key with a test quota, and answer a security questionnaire if you need one.

Contact sales Enterprise overview