Free beta · production data anonymization for AI agent evals

Synthetic case files that behave like the real ones

CasePatch anonymizes real statements, ID documents, claims and letters into realistic synthetic copies for AI agent evals, QA, vendor evaluations, training and demos. Identifying details are replaced in place, so OCR, extraction and agent workflows handle the copy the way they handle the original.

  • in-place edits: original fonts, layout and file format
  • one synthetic identity per person, across the whole case
  • every change listed for review; override any value
try it now
↑
Drop a document here PDF, PNG or JPG · up to 10 pages
Choose a file
No file handy? Try a sample:

Free: 10 pages a day without an account · 100 pages a month with a free account

statements · ID documents · claims · EOBs · letters · scans
structure preservedValues are rewritten in the original font and position. PDFs stay PDFs, so parsers and OCR read the copy like the source.
consistent across the caseEach person, account and member number maps to one synthetic value in every document.
reviewer in controlEvery change is listed by category. Override, revert or keep the original before export.
no third-party AI on uploadsThe hosted beta processes uploads with CasePatch's own engine; files are never sent to an AI provider and are deleted within 15 minutes.
Workflow

From a real case file to a usable synthetic copy

01

Upload the case

Card statements and dispute packets, KYC documents, insurance and healthcare claims, EOBs, medical records and correspondence, as PDFs, scans or photos.

02

Review each change

Detected values are listed by category next to their synthetic replacements. Compare original and copy side by side, override any value, and have it re-checked before it is applied.

03

Export

Download files with the original layout, fonts and format. The file download includes a record of reviewer overrides.

Use case: agent evals

Eval sets from real production case files

Teams building and evaluating document AI agents need test cases that look like production: real layouts, real scan quality, and the edge cases that generated test data rarely covers. CasePatch anonymizes real case files by replacing identifying details in place, so an eval set keeps the structure and difficulty of production data without carrying customer or patient identities.

Automated detection can miss values, and the output is not a formal de-identification determination. Review every page. Use test or specimen documents in the hosted beta, and run CasePatch in your own environment for production case files.

01

Real structure and edge cases

Layouts, fonts, multi-page packets and poor scans stay as they were. Only identifying details change.

02

Consistent identities across a case

The same person, account or member number maps to the same replacement in every document, so multi-document reasoning, such as matching a claimant across a claim form, an EOB and a letter, is still tested correctly.

03

Every change reviewed

Each replacement is listed with its category and can be overridden before export. The identity mapping is shown during review, so expected answers in your eval set can be updated to the replacement values.

Coverage

What we find and replace

CasePatch looks for identifying details across the data areas regulated teams handle and replaces each one in place with a realistic, format-valid value that stays consistent across the case. Every detection is listed in the review panel with its category.

PHI · HIPAA identifiers
  • Patient and member names
  • Medical record and member numbers, including Medicare beneficiary IDs
  • Provider NPI and DEA numbers
  • Dates of birth
  • Addresses and ZIP codes
  • Ages over 89

Diagnoses, medications, procedures and lab results are detected and kept by default, since evals often depend on them.

PCI cardholder data
  • Card numbers (PANs), replaced with checksum-valid numbers
  • Card last-four references
  • Expiry dates
  • Card verification codes, PINs and track data, always replaced with fresh random digits
  • Account numbers
PII
  • Names
  • SSNs and tax IDs
  • Passport, driver's license and national ID numbers
  • Addresses
  • Phone numbers and email addresses
  • Signatures
  • Faces
  • Vehicle, device and IP identifiers
Financial · KYC
  • Bank account, IBAN and routing numbers
  • ID document numbers and MRZ lines, rewritten with valid check digits
  • Organization names
  • QR codes and barcodes

CasePatch helps highlight and replace this data. It does not hold or confer HIPAA, PCI DSS or other certifications, and its output is not a compliance determination. Automated detection can miss values: review every page. Dates other than dates of birth, such as dates of service, are kept as they are. In hosted-beta uploads, faces are flagged for review rather than replaced.

Samples

Try it on public specimen documents

A card statement, an ID document, a card receipt and a Medicare notice with card. Open one to see every change, the identity mapping and the review workflow.

Redaction

Name: Susan Doe
Account: 12345-67-8907

Breaks OCR, extraction and model evaluations, and gives reviewers and vendors nothing realistic to work with.

CasePatch

Name: Nash Diaz
Account: 22244-97-8653

Plausible values in the original font, consistent across the case, so the document still parses and reads like the source.

Data handling

Uploads are deleted within 15 minutes

The hosted beta processes uploads with CasePatch's own engine; files are never sent to an AI provider. Files, results and overrides are deleted when you choose delete, when you leave the page, or 15 minutes after processing at the latest. Service logs record request types, status codes, counts and timings, not document contents or detected values.

Use test or specimen documents in the hosted beta. For production data, run CasePatch in your own environment. The hosted beta has no agent mode; in the local app it is optional and sends a case to Anthropic with your own key only when you turn it on for that case.

Run in your environment

The same engine runs on a workstation or server you control (Python 3.10+, offline after a one-time model download), so source documents stay inside your environment. Open-source release (Apache-2.0) coming soon.

Get notified

Evaluate CasePatch on your own test documents

The free beta includes:

  • 10 pages a day without an account
  • 100 pages a month with a free account
  • Personal identifiers, account and card numbers, signatures and faces
  • Reviewer overrides with leak and date checks
Create free account

limits

LimitCurrent (free beta)
File typesPDF, PNG, JPG
Pages per document10
File size20 MB
Pages without an account10 / day
Pages with a free account100 / month
Retention of uploads and resultsdeleted after 15 min, on delete, or when you leave
Faces on uploadsflagged, not replaced (replaced in samples and the local app)
FAQ

Questions

Who is CasePatch for?

Teams that need realistic documents but cannot use customer or patient data directly: card disputes, fraud and KYC operations at banks and issuers, insurance claims, healthcare claims and medical records, and legal and operations teams. Typical uses are QA and regression testing, vendor and model evaluations, staff training and product demos.

Can we use it to build eval sets for document AI agents?

Yes, that is a primary use. Anonymize a set of real case files, review each change, and use the copies as eval inputs. Because identities stay consistent across a case, tasks that depend on linking documents still have a single correct answer, which you update to the replacement values shown in review. Run CasePatch in your own environment for production case files.

Do I need an account?

No. You can try 10 pages a day without one. A free account raises that to 100 pages a month during the beta.

How do reviewers correct a replacement?

Select any change on the page or in the change list and enter a value. CasePatch re-renders it in place in the matching font and, by default, applies it to every occurrence of the original value in the case. Values that would reintroduce real data, and implausible dates, are flagged before they are applied. Each override shows the original suggestion and can be reverted.

Does CasePatch make documents compliant?

No. CasePatch produces synthetic copies and lists every change, but automated detection can miss values. Review every page before a copy leaves your team. CasePatch does not certify compliance with HIPAA, GDPR, PCI DSS or any other standard, and its output is not a de-identification determination.

Can we upload production data to the hosted beta?

Please use test or specimen documents. The hosted beta deletes uploads within 15 minutes, but it is an evaluation service. For production data, run CasePatch in your own environment.

What does it cost?

Every account gets free pages each month. Beyond that it is pay as you go with prepaid credits, from $0.05 a page and no subscription; see pricing.

Which files work best?

Text PDFs, scans and photos of printed documents up to 10 pages and 20 MB. Encrypted PDFs are not supported.

Run a sample case end to end

Free during the beta. No account or credit card required to try.

Try a sample

Sample documents and licences

  • Credit card statement: Consumer Financial Protection Bureau, “Building Blocks” sample statement. US government work, public domain.
  • Receipt photo: CORD v2 receipt dataset, Park et al. CC BY 4.0.
  • ID card photo: MIDV-500, Albanian ID specimen, Arlazarov et al. 2019. CC BY-SA 2.5; the synthetic versions shown are adaptations under the same licence.
  • Health documents: CMS “Sample Part B Medicare Summary Notice” (pages 3 and 5) and the CMS sample Medicare card. US government works, public domain. All identities are specimen data.