The initial problem
Transferring invoice information from images into records needs a consistent output: supplier, date, document number and amounts. Missing fields, duplicates and review rules matter before any accounting integration.
Technical case · Owned Optimatiza prototype
From an image to a structured record.
Capturista Digital explores how to extract invoice fields, validate their structure and add a row to an Excel workbook. This demonstration uses sample documents and is not a customer production deployment.
PythonFastAPIPydanticopenpyxlGemini
An inspectable flow before connecting real operations.
Our responsibility
The owned project covers the prototype flow and its demonstration: receiving the image, building structured data, checking the schema and writing the workbook. Messaging and email are represented in the interface.
Four stages with separate responsibilities.
Receive
Select an image from the prototype's messaging-style interface.
Extract
Gemini is the implemented extraction provider. Three synthetic samples also have predefined responses to reproduce the demonstration.
Validate
Pydantic checks the returned data structure. Type validation alone does not establish factual or tax accuracy.
Record
openpyxl adds a row to the XLSX workbook. The interface displays the record and a notification preview.
Technical decisions
Separate responsibilities
Extraction, schema validation and workbook writing are separate components. This makes their inputs, outputs and failure points easier to inspect.
Reproduce the demonstration
The three known synthetic documents have predefined fallback data. The video does not identify which extraction route produced each displayed record.
Keep review explicit
A valid schema is only one check. Representative documents, fields, duplicates and exceptions need their own review before an operational integration.
Demonstration and its limits
About the recording. This is an edited demonstration using synthetic documents. WhatsApp and the email notification are interface representations; the prototype does not prove real WhatsApp Business or email delivery.
This silent edited walkthrough uses synthetic documents. Persistent notices identify simulated interfaces; the video does not establish accuracy, latency, savings or client outcomes.
What the evidence supports
Observed and implemented
- The implementation receives an image, structures the fields and appends a row to an XLSX workbook.
- The recording provides a visual walkthrough with sample data. It does not establish extraction accuracy or processing time.
- The savings calculation multiplies document count by an assumed manual handling time. It does not subtract AI latency, review, costs or rework.
Scope limits
- No customer production operation or accounting correctness is claimed.
- Gemini uses an external AI API; the project is not described as fully local processing.
- WhatsApp Business and SMTP/Graph delivery are not proven by this prototype.
Before an operational integration
- Test representative, authorized documents and record the extraction route and errors.
- Define permissions, duplicate handling, human review and data protection.
- Connect the real channels and target systems, then measure accuracy and end-to-end times with a reproducible protocol.
Explore the related technical cases
Assess this flow against your documents.
We can review your document types, controls and target systems to define an implementation and its acceptance criteria.