Receipt AI Expense Tracker

Overview

A multimodal receipt-to-expense workflow with safe provider routing, human review, local persistence, and analytics.

Problem

Receipt images contain semi-structured financial data that must be extracted, normalized, reviewed, and stored without exposing provider credentials or silently persisting incorrect AI output.

Solution

Built a local-first Next.js application with a capability-aware AI router, structured Zod validation, Buddhist Era date normalization, editable human review, IndexedDB persistence through Dexie, and client-side spending analytics.

Outcome

The full product flow works without a database or account. AI parsing returns review data only; users explicitly confirm each receipt before it enters the local ledger and dashboard. The public deployment uses deterministic mock AI until server-side provider keys are configured.

Key Features

• Image-capability filtering prevents receipt images from being sent to text-only providers.
• Zod validation rejects impossible dates, negative totals, and malformed receipt structures.
• Reviewed records stay in the browser through a repository abstraction backed by Dexie.

Architecture

1. 01

   Receipt upload

2. 02

   Server-side AI router

3. 03

   Structured extraction

4. 04

   Zod normalization

5. 05

   Human review

6. 06

   IndexedDB repository

7. 07

   Analytics dashboard

Tech Stack

• Next.js 16
• React 19
• TypeScript
• Gemini
• Groq
• Cerebras
• Zod
• Dexie
• Recharts

Challenges & Trade-offs

• Buddhist Era dates such as 2568 normalize to Gregorian year 2025.
• AI parsing never saves automatically; persistence requires explicit user confirmation.
• Mock mode preserves the full validation and review path without spending provider quota.

Future Improvements

• Add encrypted export and import for local ledgers.
• Add duplicate receipt detection and batch upload.
• Expose field-level confidence and review history.