Skip to main content

​
Overview

Divvy’s core value proposition is bringing expense splitting to the moment of payment. Instead of manually logging expenses after the fact, Divvy integrates with Plaid to detect bank transactions automatically and prompt users to split expenses in real-time.
Implementation Status: The Plaid integration is currently in the planning/development phase. This documentation describes the intended architecture based on the project vision outlined in the README.

​
Key Features (Planned)

  • Automatic Transaction Detection: Monitor user bank accounts for new transactions
  • Smart Categorization: Use LLM-powered categorization to identify potentially shared expenses
  • Real-time Notifications: Alert users immediately when a splittable expense is detected
  • Receipt Matching: Link detected transactions to scanned receipts for itemized splitting
  • Multi-bank Support: Connect multiple bank accounts through Plaid’s unified API

​
How It Works

The Plaid integration follows this workflow:
1

Bank Account Linking

Users connect their bank accounts through Plaid Link, a secure OAuth-like flow that never exposes credentials to Divvy.
2

Transaction Monitoring

Divvy receives webhook notifications from Plaid whenever new transactions appear in the user’s account.
3

AI Categorization

Transactions are sent to Vercel Edge Functions with LLM capabilities to determine if they’re likely to be shared expenses.Example categories:
  • Restaurant/dining (high likelihood)
  • Groceries (medium likelihood)
  • Entertainment/events (high likelihood)
  • Gas/transportation (low likelihood)
  • Personal shopping (low likelihood)
4

User Notification

When a potentially shared expense is detected, Divvy sends a push notification with a deep link to quickly create a split.
5

Expense Creation

The user can:
  • Accept the suggested split and choose group members
  • Scan the receipt for itemized splitting
  • Dismiss if it’s not a shared expense

​
Architecture

┌─────────────┐
│  User's     │
│  Bank       │
└──────┬──────┘
       │
       │ OAuth Connection
       ▼
┌─────────────┐
│   Plaid     │
│   API       │
└──────┬──────┘
       │
       │ Webhooks
       ▼
┌─────────────┐      ┌──────────────┐
│  Vercel     │─────▶│  LLM API     │
│  Functions  │      │  (OpenAI)    │
└──────┬──────┘      └──────────────┘
       │
       │ Store Transaction
       ▼
┌─────────────┐
│  Supabase   │
│  Database   │
└──────┬──────┘
       │
       │ Push Notification
       ▼
┌─────────────┐
│   Divvy     │
│   Mobile    │
└─────────────┘
Plaid Link is a drop-in module that handles the OAuth flow for bank connection:

​
Android Implementation (Planned)

// Planned implementation structure
class PlaidLinkActivity : ComponentActivity() {
    
    private val linkTokenHandler = PlaidHandler.create(
        application = application,
        linkTokenConfiguration = LinkTokenConfiguration(
            token = linkToken  // Obtained from backend
        )
    )
    
    fun openPlaidLink() {
        linkTokenHandler.open()
    }
    
    private val plaidHandler = registerForActivityResult(
        OpenPlaidLink()
    ) { result ->
        when (result) {
            is LinkSuccess -> {
                // Send public_token to backend to exchange for access_token
                exchangePublicToken(result.publicToken)
            }
            is LinkExit -> {
                // User exited the flow
                handleLinkExit(result.error)
            }
        }
    }
}

​
Dependencies (To Be Added)

// Expected additions to app/build.gradle.kts
implementation("com.plaid.link:sdk-core:4.x.x")

​
Backend Components

The backend (likely a Vercel serverless function) creates a link_token: 
// Vercel Function: /api/plaid/create-link-token
import { Configuration, PlaidApi, PlaidEnvironments } from 'plaid';

const plaidClient = new PlaidApi(new Configuration({
  basePath: PlaidEnvironments.sandbox,  // or production
  baseOptions: {
    headers: {
      'PLAID-CLIENT-ID': process.env.PLAID_CLIENT_ID,
      'PLAID-SECRET': process.env.PLAID_SECRET,
    },
  },
}));

export default async function handler(req, res) {
  const response = await plaidClient.linkTokenCreate({
    user: { client_user_id: req.userId },
    client_name: 'Divvy',
    products: ['transactions'],
    country_codes: ['US'],
    language: 'en',
    webhook: 'https://your-api.vercel.app/api/plaid/webhook',
  });
  
  res.json({ link_token: response.data.link_token });
}

​
Token Exchange

// Vercel Function: /api/plaid/exchange-token
export default async function handler(req, res) {
  const { public_token, user_id } = req.body;
  
  // Exchange public_token for access_token
  const response = await plaidClient.itemPublicTokenExchange({
    public_token,
  });
  
  const accessToken = response.data.access_token;
  const itemId = response.data.item_id;
  
  // Store access_token securely in Supabase
  await supabase.from('plaid_items').insert({
    user_id,
    access_token: accessToken,  // Encrypt in production!
    item_id: itemId,
    created_at: new Date().toISOString(),
  });
  
  res.json({ success: true });
}

​
Webhook Handler

// Vercel Function: /api/plaid/webhook
export default async function handler(req, res) {
  const { webhook_type, webhook_code, item_id } = req.body;
  
  if (webhook_type === 'TRANSACTIONS') {
    if (webhook_code === 'DEFAULT_UPDATE') {
      // New transactions available
      await syncTransactions(item_id);
    }
  }
  
  res.json({ received: true });
}

async function syncTransactions(itemId: string) {
  // Get access_token for this item
  const { data } = await supabase
    .from('plaid_items')
    .select('access_token, user_id')
    .eq('item_id', itemId)
    .single();
  
  // Fetch transactions from Plaid
  const response = await plaidClient.transactionsGet({
    access_token: data.access_token,
    start_date: '2026-03-01',
    end_date: '2026-03-03',
  });
  
  // Process each transaction
  for (const transaction of response.data.transactions) {
    await categorizeAndNotify(transaction, data.user_id);
  }
}

​
LLM-Powered Categorization

As mentioned in the README, Divvy uses Vercel’s Edge Functions with LLM capabilities to categorize transactions: 
// Vercel Edge Function: /api/categorize-transaction
import { OpenAI } from 'openai';

const openai = new OpenAI({
  apiKey: process.env.OPENAI_API_KEY,
});

async function categorizeTransaction(transaction: PlaidTransaction) {
  const prompt = `
Analyze this transaction and determine if it's likely to be a shared expense:

Merchant: ${transaction.merchant_name}
Amount: $${transaction.amount}
Category: ${transaction.category.join(' > ')}
Date: ${transaction.date}

Respond in JSON format:
{
  "is_shared": boolean,
  "confidence": number (0-1),
  "suggested_category": string,
  "reasoning": string
}
`;

  const response = await openai.chat.completions.create({
    model: 'gpt-4',
    messages: [{ role: 'user', content: prompt }],
    response_format: { type: 'json_object' },
  });
  
  return JSON.parse(response.choices[0].message.content);
}

async function categorizeAndNotify(
  transaction: PlaidTransaction,
  userId: string
) {
  const analysis = await categorizeTransaction(transaction);
  
  if (analysis.is_shared && analysis.confidence > 0.7) {
    // Store transaction in Supabase
    await supabase.from('detected_transactions').insert({
      user_id: userId,
      transaction_id: transaction.transaction_id,
      merchant: transaction.merchant_name,
      amount_cents: Math.round(transaction.amount * 100),
      detected_at: new Date().toISOString(),
      category: analysis.suggested_category,
      confidence: analysis.confidence,
    });
    
    // Send push notification to user
    await sendPushNotification(userId, {
      title: `Split ${transaction.merchant_name}?`,
      body: `We detected a $${transaction.amount} expense`,
      data: {
        transaction_id: transaction.transaction_id,
        deep_link: `divvy://expense/create?tx=${transaction.transaction_id}`,
      },
    });
  }
}

​
Database Schema (Planned)

New tables to support Plaid integration:

​
plaid_items

CREATE TABLE plaid_items (
  id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
  user_id UUID REFERENCES profiles(id) ON DELETE CASCADE,
  access_token TEXT NOT NULL,  -- Encrypt in production
  item_id TEXT NOT NULL UNIQUE,
  institution_id TEXT,
  institution_name TEXT,
  created_at TIMESTAMPTZ DEFAULT NOW(),
  last_synced_at TIMESTAMPTZ
);

​
detected_transactions

CREATE TABLE detected_transactions (
  id UUID PRIMARY KEY DEFAULT uuid_generate_v4(),
  user_id UUID REFERENCES profiles(id) ON DELETE CASCADE,
  transaction_id TEXT NOT NULL UNIQUE,
  merchant TEXT,
  amount_cents BIGINT,
  detected_at TIMESTAMPTZ DEFAULT NOW(),
  category TEXT,
  confidence DECIMAL(3,2),
  status TEXT DEFAULT 'pending',  -- pending, accepted, dismissed
  linked_expense_id UUID REFERENCES expenses(id),
  
  INDEX idx_user_status (user_id, status),
  INDEX idx_detected_at (detected_at DESC)
);
Security: Access tokens must be encrypted at rest. Consider using Supabase Vault or a secrets management service like AWS Secrets Manager.

​
Environment Variables

Add to your Vercel project or backend .env: 
PLAID_CLIENT_ID=your_client_id
PLAID_SECRET=your_secret_key
PLAID_ENV=sandbox  # or production
OPENAI_API_KEY=sk-...

​
Privacy & Security

​
Data Minimization

Divvy should only request the minimum data needed:
  • Transactions product only (not Identity, Assets, etc.)
  • Last 30 days of transactions (configurable)
  • No storage of sensitive account numbers
Clearly communicate:
  • What data is being accessed
  • How it will be used
  • How to disconnect banks
  • Data retention policies

​
Encryption

  • Store Plaid access tokens encrypted
  • Use HTTPS for all API calls
  • Implement row-level security in Supabase

​
Testing

​
Plaid Sandbox

Plaid provides a sandbox environment for testing: 
// Test credentials that work in sandbox
username: user_good
password: pass_good

​
Mock Transactions

Create test transactions in sandbox: 
await plaidClient.sandboxItemFireWebhook({
  access_token: accessToken,
  webhook_code: 'DEFAULT_UPDATE',
});

​
Implementation Checklist

1

Set Up Plaid Account

  • Sign up at plaid.com
  • Get API keys (client_id and secret)
  • Enable Transactions product
2

Backend API Development

  • Create link token endpoint
  • Create token exchange endpoint
  • Implement webhook handler
  • Set up LLM categorization
3

Database Schema

  • Create plaid_items table
  • Create detected_transactions table
  • Set up RLS policies
  • Add encryption for access tokens
4

Android Integration

  • Add Plaid SDK dependency
  • Implement PlaidLink activity
  • Handle token exchange callback
  • Add deep link routing
5

Notifications

  • Set up Firebase Cloud Messaging
  • Implement push notification handler
  • Create deep link actions
6

Testing

  • Test in Plaid sandbox
  • Verify webhook delivery
  • Test categorization accuracy
  • User acceptance testing
7

Production Launch

  • Switch to production Plaid environment
  • Complete Plaid compliance questionnaire
  • Implement monitoring and error tracking
  • Deploy with feature flag

​
Cost Considerations

Plaid pricing is based on:
  • Active Items: Number of connected bank accounts
  • API Calls: Transaction syncs count toward quota
Typical costs:
  • Development (sandbox): Free
  • Production: Pay-as-you-go or monthly plans
See Plaid Pricing for current rates.

​
Alternatives to Plaid

If Plaid is not suitable:
  • Manual CSV Import: Let users upload bank statements
  • Email Parsing: Parse transaction notification emails
  • Receipt Scanning Only: Focus on OCR without bank integration
  • Teller.io: Open banking API (Europe/UK focused)
  • Finicity: Mastercard’s bank data aggregation

Build docs developers (and LLMs) love

Get started for freeTalk to us