📱 Zero-Cloud AI Budgeter

Project Overview

"Zero-Cloud AI Budgeter" is a privacy-first, local-only budgeting application designed for Android. It uniquely leverages on-device Small Language Models (SLMs) to process incoming bank push notifications, extract transaction data, and update a local, encrypted database in real-time. Our core philosophy is absolute privacy: no cloud servers, no API fees, and no third-party financial integrations like Plaid. The app offers a sleek, modern, dark-mode-first user interface inspired by premium budgeting tools like Copilot Money or Monarch Money.

Features

1. Onboarding Flow

• Privacy Pitch: Introduces the "100% Local Privacy" concept.
• Permission Gate: Guides users to grant BIND_NOTIFICATION_LISTENER_SERVICE permission.
• "Brain Download": A progress screen for downloading the local SLM model (qwen3 0.6B q4f16) to the device.
• Quick Config: Initial input for the user's primary bank to aid AI context.

2. Dashboard (Home Screen)

• "Safe to Spend": A prominent metric displaying remaining disposable income.
• Visual Charts: Interactive donut or bar charts for spending by category.
• Recent Activity Widget: A summary of the latest 3-5 AI-processed transactions (Merchant, Amount, Category).

3. Transactions Tab (The Ledger)

• Chronological List: A scrolling, date-grouped list of all transactions.
• AI Confidence Indicators: Visual cues (e.g., sparkle icon ✨) for AI-categorized transactions.
• Human-in-the-Loop (Edit Mode): Allows users to correct AI errors via a bottom sheet, ideally saving custom rules for future AI learning.

4. Budgets Tab

• Category Buckets: Progress bars showing spending against defined limits for various categories.
• Visual Warnings: Dynamic color changes (green, yellow, red) as spending approaches limits.

5. Settings & AI Control Room

• Notification Filters: Toggles to exclude notifications from specific apps (e.g., Venmo, CashApp).
• AI Hints: A text field for providing context rules (e.g., "Always categorize 'Chevron' as Gas").
• Data Management: Options to export the local SQLite database to CSV or wipe it entirely.

Technical Architecture

• Frontend: Kotlin / Jetpack Compose (Android Native)
• Backend/Logic: Entirely local to the device.
• Data Scraper: Native Android NotificationListenerService.
• AI Engine: MLC LLM for on-device inference using the qwen3 0.6B q4f16 model. This model was specifically selected for its ability to generate high-quality, structured JSON output at a small parameter count, ensuring fast and reliable parsing on mobile hardware.
• Database: Local SQLite database, encrypted with SQLCipher.

Hackathon Context

This project is being developed for the Weber State AI Hackathon (March 22 - April 3, 2026). Our goal is to demonstrate innovation, technical complexity, and a superior user experience by building a cutting-edge, local-first AI application that directly addresses the judging criteria:

• Innovation & Creativity: Eliminating external banking APIs for a truly private, notification-driven budgeting solution.
• Technical Complexity: Implementing on-device LLM inference and an encrypted local database.
• User Experience: Delivering a polished, intuitive UI inspired by leading financial apps.
• Presentation: Preparing a compelling demo video and live presentation showcasing the real-time AI capabilities.

Setup & Development

(Detailed setup instructions will follow, including environment setup, dependency management, and specific steps for both frontend and backend development.)