Documents, signatures and the paperwork layer

Collaborative editing, e-signature and OCR, built rather than bought

Collaborative document editing, a full e-signature pipeline with audit trails, and optical character recognition for the documents that arrive as photographs. Paperwork is where most business software quietly gives up and exports a PDF.

Every number here names where it came from and when it was measured.
MeasureValueSourceMeasured
Dedicated services for documents, signing and recognition4 (document service, document collaboration, e-signature service, OCR with a separate worker)kubectl get deployments -n kamo
Optical character recognition error rate0.180 CER, on our own evaluation corpusocrservice evaluation run
Known limitation we publish rather than hideSignature capture stores a flattened image, not stroke datae-signature implementation review

What building it taught us

Document work is unglamorous and it is where a great deal of real business time goes. Getting a contract signed, extracting numbers from a scanned statement, letting two people edit the same file without one of them losing an afternoon's work. Each of these is a genuinely hard engineering problem dressed as an administrative one, which is why so much software stops at 'download the PDF and email it'.

Collaborative editing means running a document server and solving multi-tenancy inside it, so that two organisations editing files of the same name never see each other's. Signing means an audit trail that would survive a dispute: who saw what, when, from where, and proof the document did not change afterwards. We learned the limits of our own signing implementation the hard way — consent is captured and the audit trail is sound, but the signature pad stores a flattened image rather than the stroke data, which matters for certain kinds of forensic challenge and which we say out loud rather than discovering in a deposition.

Optical character recognition is the piece clients most often assume is solved. It is not. Accuracy depends enormously on document quality, and a system that quotes a single accuracy figure without saying which corpus it was measured on is telling you nothing. Ours is measured, stated, and imperfect.

Built with

  • Collabora Online
  • WOPI
  • Java
  • Spring Boot
  • Python
  • Tesseract-class OCR
  • MinIO
  • PDF

What this qualifies us for

Document and workflow automationSignatures, approvals and the paperwork that eats the weekCustom web applicationsThe system your business actually runs on, built properlyAI and retrieval systemsAssistants that read your documents, and know what they are not allowed to read

Tell us what you are trying to do

A short conversation is usually enough to tell whether we are the right people. If we are not, we will say so and point you somewhere better.

Start a conversationI already use KamoCRMHow we quote