Hoopy by Michael Rork Get Pockets by HoopyGet Pockets

Case study · Pockets by Hoopy

A budgeting app built around one number: what’s safe to spend right now.

Pockets sets aside what every bill needs before it’s due, a share at a time from each paycheck, and shows what’s left over as a single number called Good to Go. I designed and built all of it, from the scheduling engine to the bank sync to the screens.

Role
Product design and engineering, solo
Built with
Vanilla JavaScript, Node.js and Express, PostgreSQL on Supabase, Plaid, Stripe
Status
Live
See it
pockets.byhoopy.com ↗

The problem

From my late teens to my mid-thirties I ran my money out of a spreadsheet built to answer one question: when do I go negative? Then Simple Bank did it properly, and I stopped needing the spreadsheet. Then Simple was sold and shut down.

I spent six years trying every budgeting app there is. They all showed me a bank balance, or a category report, or a pie chart of last month. None of them would tell me the one number I wanted: what is actually safe to spend tonight, after everything already promised to rent and bills.

Why I built it

I have ADHD. A balance that says I’m fine, while rent is already spoken for, is the exact trap my brain walks into. I needed the app to do the remembering, so I built one that does. The ADHD page on the Pockets site tells that side of it.

Home: Good to Go, what the next paycheck covers, and what’s due soon.
A bill’s Pocket fills a share at a time, so it’s full before it’s due.
Spending by category, month over month.

How I built it

  • A scheduling and allocation engine. Paychecks and bills recur weekly, every two weeks, monthly, by weekday (the third Tuesday, the last Friday) or yearly, including a leap-day bill that still lands sensibly in other years. The engine projects every pay schedule forward, finds each payday before a bill’s due date, and splits what the bill needs across them in proportion, with defined behavior for an empty window, a zero total and a window already in the past.
  • No “which paycheck pays which bill.” Plenty of budgeting apps make you assign each bill to a paycheck. I decided the user should never have to, so the engine works it out. I tried normalizing everything to annual income first and dropped it: correct on average, wrong in the days just before a due date, which is exactly when it matters. So I normalized to the paycheck instead. Every bill lands on the same scale, the amount it takes out of each paycheck, which finally shows what each expense really costs next to the others. Paid twice a month, a Costco membership is $5.42 a paycheck, not $130 a year. Netflix is $14.44 a paycheck, not $28.88 a month.
  • Rules outrank the model. Bank transaction names are gibberish, so an LLM turns them into real merchant names. Its output never wins by default. Rules carry a source and a specificity score, a person’s correction outranks a generated rule, and the most specific match wins. One correction back-fills every matching transaction in a single pass.
  • A pattern from one example. Raw bank text is full of dates, store numbers and reference codes, so a literal match only ever fires once. Pockets works out which numbers are dates by checking them against the transaction’s own date, wildcards those, keeps the real identifiers, and refuses a pattern too broad to be safe.
  • Paid calls only when someone asks. Background bank sync uses only the free endpoints. The paid ones, a live balance or a forced refresh, run only when a person taps the button.
  • Built to be tested. The parsing logic takes its database and its model as injected dependencies, so the decisions run on their own: 24 pure functions under test, with the paths that still need integration tests named in the file instead of quietly missing.
  • An AI connector. Pockets has its own MCP server behind an OAuth 2.1 sign-in, so Claude or ChatGPT can read a budget and make changes the person approves.
Every Pocket, with what it holds and when it’s due.
Connecting an AI app: a real sign-in, not a pasted key.

Stack

  • JavaScript, no framework
  • Progressive web app
  • Node.js
  • Express
  • PostgreSQL
  • Supabase auth and row-level security
  • Plaid
  • Stripe
  • Claude API
  • MCP with OAuth 2.1
  • Vercel
  • Railway

See what’s actually safe to spend, with your own paychecks and bills.

Get Pockets by Hoopy

Have something that needs building, or a role I'd fit?

michael@byhoopy.com

Burlington, Vermont · LinkedIn