Multimodal AI / local app
It started with a small flow
Upload an image, extract fields, edit them, save the receipt. The edge cases appeared quickly: Buddhist Era dates, totals that do not match line items, and personal data that should not travel farther than necessary.
What I tried
- Return a structured result from a provider or mock and run it through the same Zod validation and normalization path.
- Normalize Buddhist Era dates and send impossible dates or invalid lines back for review.
- Save only confirmed receipts in IndexedDB through a repository interface.
The flow
- 1
Receipt image → upload validation → parse endpoint
- 2
Mock / provider → Zod validation and date normalization
- 3
Human review → IndexedDB → history and dashboard
Not saving automatically became the main decision
I did not want a model response to become an expense record immediately. Confidence is not a financial guarantee, so review is part of the product flow rather than a screen to hide.
What came out of it
The synthetic fixture follows the same normalization path as provider output, supports THB and Buddhist Era dates, and routes low-confidence values back to review. That demonstrates pipeline behavior, not OCR accuracy.
What's still missing
- The public deployment uses mock AI until provider keys are configured again.
- Data stays in one browser profile with no sync, backup, authentication, or multi-device support.
- Images are stored as base64 in the local storage layer and can consume quota quickly.
What I'd change today
- Build a harder receipt fixture set with tax, subtotal, and mixed Thai formats before adding more provider routing.
- Make validation warnings easier to fix in the review screen.
- Define export and backup before keeping real receipts for the long term.
Source and demo
The repository has the complete setup, tests, and implementation details.