CyberFreezeDev
Loading
Documents read and checked automatically; only the doubtful ones reach a person.
OCR turns pixels into text. Document intelligence turns that text into fields you can trust.
OCR reads every word. Document intelligence finds the fields you need, checks them, and sends failures to a person with the image.
Plain OCR gives you every word on the page in reading order. That is useful for search, but your accounts team does not need every word. They need the vendor's GSTIN, the invoice number, the date, the taxable value, CGST, SGST or IGST, and the total, in the right columns.
Document intelligence is the layer that finds those fields, whatever the layout, and then checks them. Does the GSTIN have 15 characters in the right pattern? Do the line items add up to the taxable value? Is this the same invoice number from the same vendor you already booked last week? A field that fails a check, or that the model was unsure about, goes to a person with the image next to it.
The extracted data then feeds the rest of your process: bills into accounts (our GST and accounting automation work picks up from here), KYC details into onboarding, or claim forms into an approval workflow.
| Document | Fields we extract | Checks before it is accepted |
|---|---|---|
| GST invoices and bills | Vendor, GSTIN, invoice number and date, HSN or SAC, taxable value, tax, total | GSTIN pattern, line totals, CGST and SGST versus IGST, duplicate invoice numbers |
| PAN cards | Name, PAN, date of birth | Ten-character PAN pattern, name matched to the application |
| Aadhaar and other IDs | Name, ID number, address | Number format; Aadhaar masked to its last four digits before storage |
| Bank statements | Date, narration, debit, credit, balance | Running balance reconciles line by line |
| Application and claim forms | Fields specific to your form, ticks, signatures present | Required fields filled, values within allowed ranges |
| Prescriptions and lab reports | Patient name, date, items listed | Always sent for human review; never auto-accepted |
Handwriting, poor photos and stamps over text lower accuracy. We measure it on your own documents before promising anything.
Past documents as they really arrive: good scans, bad photos, multi-page PDFs.
You share a set of past documents that represents what actually arrives: good scans, bad phone photos, multi-page PDFs, different vendors. We sign an NDA first if you want one, and personal data stays on systems you approve.
Agreed fields, formats and rules become the test set every version is scored on.
We agree the exact fields, formats and rules with the people who use the data. That labelled sample becomes the test set every version of the pipeline is scored against.
Text extraction, OCR or a vision model per document type, chosen by measurement.
Clean digital PDFs often need no OCR at all, just text extraction. Scans and photos go through an OCR engine or a vision language model. We pick per document type by measuring on your sample, not by preference.
Validation, duplicate checks and a side-by-side screen for correcting and approving.
Validation rules, duplicate detection and a review screen where a person sees the image and the extracted fields side by side, corrects them and approves.
Approved data written to your systems, with a record of every correction.
Approved data is written to your sheet, database or accounting system. Every document keeps a record of what was read, what was corrected and by whom.
Dozens or hundreds of documents a day, in known types, keyed in by hand.
Dozens or hundreds of documents a day in a handful of known types, where someone currently keys the data in. CA firms handling client bills, clinics digitising intake forms, insurers and lenders collecting KYC, and logistics teams reading delivery paperwork.
A few documents a week, or handwritten archives that only need search.
A few documents a week. A good template and a careful person is cheaper. Also archives of mostly handwritten notes where nobody needs structured fields, only search.
IDs, medical and financial records processed on your own server with open-weight models.
ID cards, medical records and financial statements can be processed entirely on your own server with open-weight models, so nothing is sent to an outside API. See custom AI systems for how we deploy private models.
Documents picked up from a mailbox or WhatsApp, with a request for clearer photos.
Many documents arrive as chat photos or attachments. The pipeline can pick them up from a mailbox or a WhatsApp number and reply when a clearer photo is needed.
Described without the client's name, as agreed.
An OCR pipeline for invoices, IDs and forms where failed or low-confidence fields go to a reviewer before anything moves on.
We built an OCR pipeline for invoices, ID documents and forms with human review. Documents are read, fields are checked against format and arithmetic rules, and anything that fails or falls below the confidence threshold is queued for a reviewer who sees the original image beside the extracted values. Only approved records move downstream.
OCR and document AI is also one of the research tracks at CyberFreeze Labs, where we test new models against the same kinds of Indian documents before they reach client work.
It depends on your documents, so we do not quote a number before measuring. We score the pipeline on your labelled sample and show you field by field where it is reliable and where it needs review. Low-confidence fields always go to a person.
Printed Tamil, Hindi and other Indian scripts can be read by current OCR engines and vision models with varying quality; handwriting is harder. We test on your samples and route handwritten fields to review by default.
Aadhaar use is regulated by UIDAI. Our default is to mask all but the last four digits before anything is saved, and to keep the original image only if you have a lawful reason to. Your compliance team or advisor should confirm what you may keep.
Yes. We build integrations with Tally so approved purchase bills can be posted as vouchers instead of being retyped. A person still approves the entry.
Only if you choose a hosted model whose servers are elsewhere. We can run the whole pipeline on a server in India or in your own office if the data must stay local.
Last updated