📒 KhaataKitab

Observe • Understand • Verify • Assist
Turning raw transaction signals into intelligent financial records

Live App

The live deployment is primarily intended for demonstration. Some capabilities — particularly native Android SMS ingestion and cloud synchronization — remain under active development.

Most finance apps expect users to input everything manually.

KhaataKitab behaves more like an assistant that watches, processes, and verifies — reading signals that already exist (bank SMS, receipts) and turning them into structured, categorized ledger entries, with a human in the loop to confirm or correct.

flowchart TB
    U[User]

    U --> WEB[React Web App]
    U --> AND[Capacitor Android App]

    WEB --> AUTH[Supabase Auth]
    AND --> AUTH

    WEB --> DB[Dexie / IndexedDB]
    AND --> DB

    AND --> SMS[SMS Ingestion]
    SMS --> PARSER[Transaction Parser]

    PARSER --> CLASSIFIER[Local Categorization Engine]
    DB --> CLASSIFIER

    CLASSIFIER --> REVIEW[Human Verification]
    REVIEW --> DB

    CLASSIFIER --> AI[Supabase Edge Functions]
    AI --> LLM[AI Provider]

    DB --> INSIGHTS[Insights Engine]
    AI --> INSIGHTS

flowchart LR
    A[Transaction Signals] --> B[Parse]
    B --> C[Categorize]
    C --> D{Confidence}
    D -->|High| E[Ledger]
    D -->|Low| F[AI Assistance]
    F --> G[User Verification]
    G --> H[Learn From Correction]
    H --> C
    E --> I[Insights]

• Observe — capture transaction signals from SMS and receipts
• Understand — extract and categorize transaction data
• Verify — confidence-based review and human confirmation
• Assist — generate summaries and financial insights
• Learn — incorporate user corrections into local categorization

• Regex-based parser extracts amount, direction (credit/debit), payment method, last 4 digits, and reference ID from raw bank SMS text
• Runs entirely client-side — no data leaves the device for parsing
• Native Android auto-read (real-time SMS listener) is in progress; current build supports import of sample SMS for testing the parsing pipeline

A cascading pipeline that only escalates cost when it needs to:

1. User-learned mappings (IndexedDB) — instant, from your own corrections
2. Merchant dictionary match — e.g. "Swiggy" → Food & Dining
3. On-device Naive Bayes classifier — probabilistic categorization with confidence scoring, Laplace smoothing, online learning from corrections
4. AI fallback (Supabase Edge Function → AI provider) — triggered only when local confidence is low

• Transactions below a confidence threshold are flagged Needs Review
• Correcting a category updates the on-device classifier's word-frequency table, improving it over time
• Manual entries are cross-checked against parsed SMS (amount/time/merchant/method) to support verified vs. needs-review status

• Monthly income/expense summaries, category breakdown, cashflow trend charts
• Financial Health Score (0–100) from three weighted factors: income regularity, expense control, consistency
• Cashflow "prediction" is currently a deterministic heuristic (rolling average), not a trained forecasting model — labeled accordingly in-app, because the distinction between a heuristic and a model matters

• CRUD for stock items with quantity and computed value
• Ledger ↔ Inventory integration (in progress): sales transactions will automatically decrement stock

• Fully responsive: 1 column (mobile) → 2 (tablet) → 3 (desktop)
• Offline-first: reads and writes continue uninterrupted with no network connection, backed by IndexedDB (Dexie)
• Dark mode, smooth transitions, real-time reactive updates on data change

Frontend: React · TypeScript · Vite · Tailwind CSS · shadcn/ui Data & Persistence: Dexie · IndexedDB Backend: Supabase · PostgreSQL · Edge Functions AI: AI Edge Functions · local Naive Bayes classifier Mobile: Capacitor · Android Visualization: Recharts

• Multi-tier categorization pipeline balancing cost, latency, and accuracy (local-first, cloud fallback only when needed)
• Online-learning classifier that improves from user corrections without a server round-trip
• Offline-first architecture with reactive local persistence
• Structured extraction from unstructured bank SMS text via regex parsing
• Real authentication via Supabase Auth, replacing the earlier local-session prototype

git clone <YOUR_GIT_URL>
cd <PROJECT_NAME>
npm install
npm run dev

Environment variables required (see .env.example):

VITE_SUPABASE_URL=
VITE_SUPABASE_ANON_KEY=

• 👥 Contacts / party ledger ("who owes me, whom do I owe") — the core khata use case, in progress
• ☁️ Genuine cloud sync with conflict resolution
• 📦 Ledger ↔ inventory integration (sales auto-adjust stock)
• 📱 Native Android SMS auto-capture
• 🧪 Automated test coverage (parser, classifier)
• 💳 Credit/trust scoring for repeat customers
• 🌍 Multi-language input (Hindi/Marathi voice and text)

The future of apps is not interaction. It's automation with intelligence — and a human who can still check its work.

MIT License