Building a MoneyForward-Beating AI Budget Advisor in Flutter Web (Supabase + Claude)

Why Build This

Amazon just launched "Rufus Buy for Me" — an AI that browses and purchases for you. The counter-move for a personal finance app? Make the AI work for your wallet, not against it.

I'm building 自分株式会社, an AI life management app that replaces 21 SaaS tools (Notion, Evernote, MoneyForward, Slack, etc.) with one integrated platform. Today's task: transform a 128-line stub budget page into a full AI-powered financial advisor.

What We Built

A 4-tab budget page:

1. Overview — Income/expense KPIs + top 5 spending categories
2. Budget — 15-category budget vs. actual with progress bars
3. AI Savings Advice — Claude analyzes spending and generates 3 actionable tips
4. Future Simulator — Compound interest calculator for wealth projection

The AI Savings Advisor

Instead of creating a new Edge Function, we reused the existing ai-assistant EF — just with a new prompt:

Future<void> _fetchAiAdvice() async {
  final prompt = '''
Analyze this month's ($_selectedMonth) household data and suggest 3 specific saving tips.

Total income: ¥${_fmt.format(totalIncome)}
Total expenses: ¥${_fmt.format(totalExpense)}
Spending by category:
$breakdown

Provide exactly 3 practical, actionable saving suggestions.
''';

  final response = await _supabase.functions.invoke(
    'ai-assistant',
    body: {'action': 'chat', 'message': prompt},
  );
}

Prompt engineering notes:

• Pass structured monthly data (not raw transactions — the AI doesn't need every line item)
• Constrain to "exactly 3 tips" to keep responses concise and actionable
• Ask for "practical" suggestions to avoid generic advice like "spend less"

The ai-assistant EF routes this to Claude Sonnet and returns suggestions like:

• "Your dining out (¥45,000) is 23% over budget — try batch cooking 3 meals/week to cut this by ¥15,000"
• "Switch your streaming subscriptions to annual billing — saves ~¥8,400/year"
• "Your transport costs suggest frequent taxis — a monthly pass would save ¥12,000"

Compound Interest Simulator

Pure Dart implementation (no packages needed):

void _runSimulation() {
  final principal = double.tryParse(_initialAmountCtrl.text) ?? 0;
  final monthly = double.tryParse(_monthlyAddCtrl.text) ?? 0;
  final rate = (double.tryParse(_returnRateCtrl.text) ?? 0) / 100 / 12;
  final months = (int.tryParse(_yearsCtrl.text) ?? 0) * 12;

  double result = principal;
  for (int i = 0; i < months; i++) {
    result = result * (1 + rate) + monthly;
  }
  setState(() => _simResult = result);
}

Example: ¥1M initial + ¥30K/month + 5% annual return + 20 years = ¥15.4M
Principal contributed: ¥8.2M → Interest earned: ¥7.2M — compound interest does the heavy lifting.

Edge Function Pattern: Reuse app_analytics for Multi-Purpose Storage

Rather than creating a dedicated budget_data table, we use app_analytics with a source column as a discriminator:

// In budget-financial-planner EF
await supabase.from('app_analytics').upsert({
  user_id: userId,
  source: 'budget',           // discriminator
  category: budgetCategory,
  amount: budgetAmount,
  month: targetMonth,
  metadata: { breakdown: categories }
});

Why this works: The app_analytics table already has user_id, category, amount, metadata columns. For budget data, we just add source: 'budget' and query with .eq('source', 'budget'). No new migration needed.

This pattern scales well when you have many feature-specific datasets that don't need complex relational queries.

Competing with MoneyForward

MoneyForward's strengths: bank account sync, automatic categorization, long history.

Our approach: AI-first, not data-first. Instead of syncing every transaction, users input monthly summaries and the AI does the analysis. Less friction, more insight.

What we added that MoneyForward doesn't have:

• Natural language AI advice (not just charts)
• Integrated with the rest of your life data (tasks, goals, habits)
• Future wealth projection tied to your actual budget

Key Takeaways

1. Reuse existing Edge Functions — adding a new prompt is cheaper than a new EF
2. Constrain your AI prompts — "exactly 3 tips" beats "give me advice"
3. Generic tables > specific tables — app_analytics with a source column handles dozens of use cases
4. Compound interest is powerful — build a simulator and show users the math

Try it: 自分株式会社

buildinpublic #Flutter #Supabase #AI #fintech #personalfinance