Personal Finance App

On-device AI can handle sensitive financial data without compromising privacy. Building in public to explore the future of private, intelligent personal finance.

Tech Stack
  • Swift
  • SwiftUI
  • CoreML
  • On-device LLM
  • Vision Framework

  • PROJECT TYPE

    Personal Project

  • TIMELINE

    TBD

  • STATUS

    Building in Public
  • YouTube Series Coming Soon
Personal Finance App - Building in Public

The Problem

Personal finance apps require trusting cloud services with your most sensitive data—bank balances, spending patterns, income details. For many users, that's a dealbreaker. They'd rather track finances manually (or not at all) than upload their financial life to someone else's servers.

The trust problem isn't just about security breaches. It's about control. Who has access to your data? How is it being used? Can you ever truly delete it? These questions create friction that prevents adoption, even when users acknowledge they need better financial tracking.

Meanwhile, manual tracking is tedious. Typing transactions is time-consuming. Categorizing expenses is repetitive. Users want intelligent insights—budgets, trends, recommendations—but not at the cost of privacy.

The Hypothesis

I hypothesize that on-device AI can deliver intelligent personal finance features without cloud dependency. All processing happens locally on the user's iPhone—LLM runs on-device, data never leaves the device, AI-powered insights without server roundtrips.

This approach solves both the privacy problem (complete data ownership) and the speed problem (no network latency for common tasks). The trade-off is model capability—on-device LLMs are smaller and less powerful than cloud models. But for personal finance, do you need GPT-4 level reasoning, or is a smaller, faster model sufficient?

Key Decisions

  • On-device processing—privacy by architecture: No cloud servers, no API calls, no data uploads. Financial data stays on the device, encrypted at rest using iOS security features.
  • Photo-first data capture: Take a picture of a receipt, and Vision Framework + LLM extract the transaction details. Minimal typing required.
  • Proactive insights and reminders: The LLM analyzes spending patterns and surfaces insights—"You've spent 30% more on dining this month" or "Rent payment due in 2 days."
  • CoreML for on-device LLM inference: Leveraging Apple's CoreML framework to run language models directly on iPhone's Neural Engine—no internet required.
  • Building in public: Documenting the entire build process on YouTube to show how product decisions, technical trade-offs, and user testing shape the final product.

What I'm Learning (In Progress)

This project is currently in development. Follow along on YouTube (coming soon) to see:

  • How to evaluate on-device LLM capabilities vs. cloud models
  • Trade-offs between model size, inference speed, and feature quality
  • Product decisions around what features need AI vs. traditional logic
  • User testing insights on privacy-first finance tools
  • Technical challenges of running LLMs locally on mobile devices

PM Skill Being Developed

On-Device AI & Privacy-First Design: This project is teaching me how to design AI products where privacy isn't a feature—it's the foundation. It's forcing me to think about capability constraints (what can smaller models do well?) and user expectations (when do users want AI help vs. manual control?).

Building in Public: Documenting the journey publicly creates accountability and demonstrates product thinking in real-time—not the polished case study, but the messy process of building, testing, and iterating.

Back to AI Lab