Side Project AI-Built

Budget
Tracker

A PWA built with AI — from idea to shipped product in 2 weeks.

Budget tracker home screen
The Problem

Every budget app got in the way.

  • Friction

    Logging a transaction was buried 4+ taps deep. By the time you got there, the motivation was gone.

  • Privacy

    Most apps want bank credentials. Handing over access to a third-party service never felt right to us.

  • The Bar

    Open the app → log the transaction → done. That was it. Nothing more.

Process

How I built it

No wireframes on a blank canvas. I used Design OS to structure the build before writing a line of code.

  • Design OS

    Defined the product scope, data model, and component structure using Design OS — a tool that separates "what to build" from "how to build it" before any code is written.

  • Brainstorm

    Used the brainstorming skill to pressure-test requirements: what's the minimum viable feature set? What can be cut? What order do things need to be built in?

  • Spec → Code

    Design OS exported a complete handoff — data types, component specs, screen designs. That spec went straight to Claude Code as context.

  • Shipped

    A fully working PWA in 2 weeks. Not a prototype — a real app my wife and I use daily for household expenses.

The App

What it does

Log fast, switch easily

  • Home screen opens straight to logging
  • Natural language: "Spent 500 on lunch"
  • One tap to switch personal ↔ joint account
Home screen with natural language input Joint account view

History, insights, AI

  • Monthly breakdowns with opening/closing balances
  • Category spend dashboard at a glance
  • Groq-powered chat: "What did I spend on food this month?"
Transaction history Insights dashboard

Private by default

  • No bank connections, no accounts, no servers
  • All data in local storage. Nothing leaves the device
  • Export any date range as a file
Export with date range selection Settings and privacy controls
Reflection

What I learned

  • Design first

    Using Design OS before touching code meant every AI prompt had real context. The output was sharper because the input was structured.

  • Be the user

    Solving your own problem is the fastest feedback loop. If something felt off during daily use, I fixed it the same day.

  • Ship it

    A designer who can spec, build, and ship with AI is a different kind of designer. This project proved that to me.