Main project
La Nueva Goleta Administración
I started it to keep the restaurant's invoices, prices, cash and expenses in one place. It runs in a pilot environment today and gets better with every round of testing.
Context
I work in the admin side of La Nueva Goleta. Entering invoices by hand, tracking prices, logging cash and expenses: all of it ate hours and ended up scattered across different places.
Problem
A good part of the admin was done by hand and the information ended up scattered. That made it harder to keep track of prices, expenses and suppliers.
Product decisions
The most important decision was not to automate blindly. OCR proposes and a person confirms: every line can be reviewed and corrected before it enters the system.
- Nobody accepts an extraction without reviewing it.
- The original document always stays next to the data.
- Reprocessing a document cannot wipe out work already reviewed.
- Separate profiles for owner, administrator, employee and accountant.
Solution
I built a system that brings those processes together, keeps a history, and lets you review every piece of data before approving it.
Verified features
- Invoice upload and storage
- Information extraction via OCR
- Reading products by their position inside the document
- Review and correction of extracted lines
- Safe reprocessing of documents
- Alias and equivalence memory per supplier
- Matching with high, medium or low confidence
- Normalisation per litre, kilogram or unit
- Price change alerts
- Daily cash management
- Expense and payment records
- Main dashboard and reports
- History and audit trail
- Independent pilot environment
- Responsive design
- Groundwork for owner, administrator, employee and accountant profiles
Technical challenges
- Every supplier sends invoices in a different format, so a single template is useless.
- The same product is written differently depending on the document, so I store aliases per supplier.
- To compare prices you have to normalise units: a price per case is not a price per litre.
- Matching has to allow for doubt, so it returns a confidence level instead of a yes or no.
Validation
I test the system with everyday documents and situations in a pilot environment kept apart from daily use, so a mistake of mine never touches the restaurant's data.
The screenshots on this page come from that pilot environment, with demo data.
Learnings
- Knowing the operation from the inside changes what you build: half of these features would never have occurred to me from outside.
- An automation nobody trusts does not get used. Keeping the review visible is what got it adopted.
- AI speeds up my development, but deciding what to build and what to leave out is still my job.
Next steps
- Extend the pilot to more document types and more suppliers.
- Finish the separate access profiles.
- Measure how much time it saves, using the system's own data.
