OCR and document intelligence
Zapper Edge reads scanned, photographed and handwritten documents, keeps the layout, and returns structured output with every page accounted for, all inside your harness controls.
Digitize scanned and handwritten documents at scale into page-accurate text, tables and structured data that your systems and AI models can use.
- Vision-language OCR engine
- Layout and table detection
- Handwriting and scan support
- GPU acceleration
- Page-marked Markdown
- JSON block list
- HTML tables
- PDF and image input
- Browser-based batch console
- REST API
- Command line
- Folder browsing
The challenge
Most enterprise knowledge still sits in scans: merged PDFs, photographed forms, signed letters, tables and handwritten notes. People retype them or leave them unread. Search cannot find what a machine cannot read, and AI models cannot reason over a page they never received.
Basic OCR makes it worse in quiet ways. Tables collapse into lines of text, columns merge, stamps and card scans come back empty, and page numbers drift from the original file. Downstream teams then trust output that nobody can trace back to the source page.
What it does
Scanned and handwritten input
Reads scanned pages, photographs and handwriting from PDFs and images, including card scans and pictures that contain text, so content locked in paper-era files becomes machine readable.
Layout and table preservation
Detects headings, paragraphs, lists, captions, footnotes and tables as separate blocks in reading order, so columns and tables keep their meaning instead of flattening into noise.
Page-accurate output
Every page appears in the output with its own marker, including blank and image-only pages, so page numbers always match the source PDF and any value can be traced back.
Structured data extraction
Delivers text as Markdown and a full block list as JSON with position and table data, ready for field extraction, search indexing, analytics or retrieval by an AI model.
Batch processing at scale
Runs whole folders of documents one after another with chosen concurrency, skips files already done, and records a one-line reason for any failure without stopping the batch.
Safe, repeatable runs
Originals are never changed. Results are written atomically beside each source file, empty parses are treated as failures, and a stopped run resumes where it left off.
How it works
Point to the documents
Choose a folder of PDFs or images, or send a file to the REST API. The source files are read only and are never moved, renamed or rewritten.
Read each page
A document parsing engine reads every page, including scans and handwriting, and identifies text, tables, lists, captions and pictures along with their position and reading order.
Build structured output
Blocks are assembled into page-marked Markdown plus a lossless JSON block list. A header records the engine, date, page count and a fingerprint of the source file.
Hand over and verify
Output sits beside each source file under the same name for search, extraction, classification or quality checks. Failed files are listed with a reason, ready to retry.
Under the hood
- Document reading
- Vision-language OCR engine
- Layout and table detection
- Handwriting and scan support
- GPU acceleration
- Output formats
- Page-marked Markdown
- JSON block list
- HTML tables
- PDF and image input
- Interfaces
- Browser-based batch console
- REST API
- Command line
- Folder browsing
- Runtime
- Docker containers
- NVIDIA GPU host
- Private deployment
Business outcome
Paper archives become searchable. Scans and photographs turn into text that search, analytics and AI models can use, so knowledge held in old files stops being invisible to the business.
Less manual re-keying. Structured text and tables replace retyping from images, freeing teams from data entry and cutting the transcription errors that come with it.
Output you can trace. Page-accurate results with source fingerprints let reviewers confirm any value against the original page, which builds trust in what downstream systems consume.
A foundation for AI use. Clean, structured document text gives retrieval, extraction and classification workloads reliable input, so AI adoption starts from readable data rather than guesses.
Where it is used
- Digitizing archives of scanned contracts, letters and forms so they can be searched and reviewed.
- Extracting tables and fields from invoices, statements and reports for finance and operations teams.
- Reading identity cards, certificates and handwritten forms collected during onboarding or verification.
- Preparing document collections as clean text for retrieval, analytics or model-based extraction.
Frequently asked questions
Can it read handwriting and poor scans?
It is built to read scanned pages, photographs and handwriting, and it keeps tables and layout. Results depend on scan quality, which is why a separate quality check can grade pages before reading.
What does the output look like?
Each document produces page-marked Markdown for reading and search, plus a JSON block list with block types, positions and table data. Both are written beside the source file under the same name.
Does it change or move my original files?
No. Source documents are only read. Output is written next to them, existing results are skipped unless you choose to overwrite, and a failed file leaves nothing half-written behind.
Where does processing run, and how does it scale?
It runs as containers on your own infrastructure with an NVIDIA GPU for reading. Batches process folders of documents with a chosen concurrency, and can be started from a browser console or the API.
See it on your own documents and data
Tell us what you want to automate and we will show you how it runs inside the harness.