Grocery Budget

Expo React Native

Budgeting and meal-planning app: auto-budget from bank transactions (Plaid), receipt parsing with grocery subcategory splits, recipes with cost-per-serving, and AI-generated shopping lists.

React Native + ExpoTypeScriptOpenAIConvexClerkPlaid

View the Code

React Native + Expo on GitHub

View on GitHub

Problem & Context

Grocery spending is hard to track and plan. Manual budgeting is tedious; bank feeds don't tell you what you bought or how it maps to meals. Receipts pile up. Meal planning and shopping lists live in separate tools.

Grocery Budget unifies envelope-style budgeting (with Plaid auto-import), receipt capture and itemization, recipes with cost-per-serving, weekly meal planning, and AI-generated shopping lists. Receipts can link to bank transactions and split grocery spend into subcategories (e.g., Costco vs household vs eating out).

Core Pillars

Auto-budgeting from bank transactions — Plaid integration with envelope-style budgets, parent + subcategory hierarchy, transfer tracking, and rule/AI-assisted categorization.
Receipt → items + grocery subcategories — Camera capture, AI receipt parsing, item-level categorization, and linking receipts to transactions to split grocery spend.
Recipe → ingredients → cost-per-serving — URL ingestion, manual entry, saved recipes, weekly meal plan with slots, and AI-generated shopping list with pricing.

Categorization Pipeline

Transactions are categorized by strict priority: user rules first, deterministic mappings second, AI only as last resort.

1. User rule cache
Fuzzy merchant match — wins over everything
2. Deterministic mapping
MCC codes, keywords
3. AI fallback
Only when unknown merchant and no MCC mapping

App Structure

Dashboard — Spend vs budget, alerts, sync/scan quick actions

Budget — Period selector, allocations, rollover config

Transactions — List, filters, category picker, receipt link suggestions

Receipts — Needs review, linked/unlinked, edit parsed fields

Meals — Meal plan, recipes, shopping list

Settings — Account, plan, notifications, household

AI System (Paid Tier)

Every AI output stores confidence scores, sources (receipt vs web), and rationale. Low-confidence fields require review. User edits feed back into rules and caches.

Transaction categorization fallback
Receipt OCR + parse (image → structured receipt)
Item normalization (raw → canonical)
Recipe URL ingestion
Recipe cost-per-serving estimation
Shopping list generation from meal plan

Key Decisions

Guest mode with local-only storage

Anonymous users get full manual functionality via expo-sqlite (budgets, transactions, recipes, meal plan) without Convex or Plaid. On sign-in, merge flow uploads local data with local-wins conflict resolution.

Tradeoff: No offline write queue in MVP; signed-in users require connectivity for writes.

Categorization priority: rules > deterministic > AI

User merchant rules (fuzzy match) always win. MCC/keyword mapping next. AI only for unknown merchants with no mapping. Manual categorization auto-creates fuzzy rules for future transactions.

Tradeoff: AI calls are minimized and reserved for paid tier, reducing cost and latency.

Receipt linking with mismatch tolerance

Link by time, store, and total. Within 5% tolerance allows link but flags for review. Linked receipts split grocery transactions into subcategory allocations; unlinked receipts do not affect budgets.

Tradeoff: Some false matches possible; review UX handles flagged cases.

Cost-per-serving from receipt history first

Pricing priority: (1) most recent purchase within 30 days from receipts, (2) AI web search. Free tier: user selects match; paid: AI suggests, flags uncertain for review.

Tradeoff: Accuracy depends on receipt history quality; newer users get more web-sourced estimates.

Screenshots

What I'd Improve

Offline write queue for signed-in users
Pantry tracking with auto-consumption from recipes
Benchmarking (50th percentile spend) using anonymized opt-in data
Store routing optimization for shopping lists
Data export