Chooses the right reading method
Text PDFs are read directly, while scanned PDFs and images use the image-processing path.
marvinjb.devLIVE AI SYSTEM · EXTRACTION AGENT
Turn unstructured invoices into validated data.
Upload a PDF or image of an invoice. The AI extracts the vendor, dates, totals, and line items, validates the result, and lets you ask questions about the document.
LIVEPUBLIC DEMO
01 · LIVE DEMO
Upload an invoice and see what the AI extracts
Upload a PDF, scanned PDF, JPG, or PNG. The system reads the document, extracts the invoice fields, validates the result, and lets you inspect the structured data or ask questions about the invoice.
- 01Upload invoice
- 02Extract data
- 03Review Table or JSON
- 04Ask questions
PDF · JPG · PNG / 5 MiB MAX
Drag and drop your invoice here.
PDF, scanned PDF, JPG, or PNG. One invoice at a time.
Upload → AI Extracts → Explore → Ask
01PDF & image inputAccepts PDF, JPG, and PNG invoices
02Validated structured dataExtracts invoice fields and checks the result
03Invoice Q&ALets users ask questions about the extracted document
04Production deployedFastAPI · Docker · Nginx
02 · HOW IT WORKS
The AI chooses the right way to read each invoice.
Text-based PDFs are read directly. Scanned PDFs and images are processed visually. Both paths produce the same validated invoice data.
INPUTPDF or image
→DETECTDetect document type
→READRead text or image
→EXTRACTExtract invoice details
→VALIDATEValidate the result
→OUTPUTInvoice data + Q&A
03 · ENGINEERING DECISIONS
How the extraction stays consistent
The system can read different invoice formats, but every document is converted into the same validated invoice structure before it is returned.
Returns the same invoice format
Every supported document ends as the same structured invoice result.
AI returns structured data
The model returns specific invoice fields instead of free-form text.
Invoice Q&A is separate
Extracting the invoice and asking questions about it are handled as separate requests.
04 · RELIABILITY
What happens when something goes wrong
The system checks files before processing them, validates the AI result before showing it, and gives the user a clear error when something fails.
Checks files before upload
Empty, unsupported, or oversized files are rejected.
Checks AI results
Invalid extraction or Q&A responses are not shown as valid data.
Shows clear errors
Upload, extraction, timeout, network, and question failures stay visible to the user.
Tested for regressions
File handling, API behavior, document routing, and structured outputs are covered by tests.
05 · TECHNOLOGY STACK
FrontendReact · TypeScript
BackendPython · FastAPI · Pydantic · pypdf
AI & Document ProcessingOpenAI Structured Outputs · multimodal vision
DeploymentDocker · Nginx · Ubuntu VPS · HTTPS · health checks
INSPECT THE IMPLEMENTATION
See how it was built
Explore the API flow, validation, tests, and deployment setup behind the Extraction Agent.