Skip to content

work / keep-receipt · Live

Keep Receipt

Snap a receipt, AI reads it, your spending explains itself.

Week of Claude Maxing

I needed a digital record of business expenses for tax time, and a clean line between personal spending and the businesses. Budgeting insight came along for free once everything was captured. The design constraint was cost: read the receipt on-device with OCR first, and only fall back to Gemini vision when the scan is genuinely unreadable, so the AI bill stays near zero.

Why I built it

Shoebox receipts and no idea where the money goes, especially across personal and business.

The problem

Photograph a receipt and on-device OCR reads it, with an AI vision fallback only on low confidence. Auto-categorised spend, personal-versus-business tagging, insights, and an Australian tax lens with entity tracking.

What it does
  • React 19
  • Dexie
  • Tesseract.js
  • Vision LLM
  • PWA
Stack
Cheap local first, expensive AI only on low confidence.
What I learned

Screenshots from the app in actual use, not mockups.

Keep Receipt finance dashboard for FY 2026 to 27, showing spend, estimated deductions, receipt count and a per-account breakdown.
A scanned receipt in Keep Receipt with the merchant, total and date read automatically from the photo, ready to categorise and tag.
The desktop Finance view of Keep Receipt, adding a tracked subscription with its billing cycle and a cancel reminder.

The real thing

A pure-frontend replica with mock data: nothing is saved, nothing leaves the page. The real app is richer; this is the feel of it.

loading the demo…

Play with it

The thinking

user · pains · gap · tools

Australians juggling personal and business spend (employees, sole traders), starting with me.

Target user

  • Shoebox receipts and lost deductions at tax time
  • Manual categorising never happens
  • Cloud finance apps want your data on their servers

Pain points

A privacy-first, ATO-aware receipt keeper: on-device OCR by default, AI vision only when needed, multi-entity tax lens built for Australian FY rules.

Gap in the market

  • Tesseract.js OCR
  • Kimi vision (fallback)
  • Claude Code (build)
  • Dexie
  • Google Drive backup

AI + tooling

  • Capture at the point of action: snap now, categorise never
  • Loss aversion: surfacing the deductions you are about to miss
  • ATO myDeductions parity model with multi-entity tagging
  • Privacy by design: hybrid cheap-local OCR, AI only on low confidence

Frameworks

  • Live and used across several real expense runs
  • Hybrid OCR-first pipeline keeps the AI vision cost to genuine fallbacks only
  • TBD: receipts captured, deduction totals surfaced

Success metrics

Real-time voice lessons that feed a spaced-repetition word bank.

Next: Mandarin Tutor

← all work