Budget G.P.A.: Case Study

Background

The Verdict

Budgeting apps show you data. Almost none of them render a verdict.

1Number a person can act on: a GPA on the familiar 4.0 scale

$39.99Per year for Pro, against $109 for YNAB and $99.99 for Monarch

5Pillars, so the grade can say which part of your month cost you

At work I grade enterprise cloud subscriptions with a weighted scorecard, and the effect is immediate: a letter grade ends the debate about whether things are fine. Personal finance apps have the opposite problem: endless charts, categories and balances, but no answer to the only question people have, “am I doing okay?”

So the grade is the product. Every month you get a report card: a letter grade, a GPA, and the top fixes that would raise it. The competition competes on feature breadth; this app competes on a single legible number.

The Method

One Month’s Report Card

Five pillar scores, each from 0 to 100, weighted into one grade. The numbers below are illustrative, run through the app’s real weights.

Pillar	Weight	What It Measures	Example Score	Points
Budget Adherence	35%	How closely actual spending matches each category’s budget	88	30.8
Financial Safety	25%	Emergency fund, debt to income, and cash left at month end	80	20.0
Spending Stability	15%	Whether spending holds steady month to month, in dollars	90	13.5
Needs vs Wants	15%	How close the split is to a 50/30/20 framework or your own targets	85	12.8
Forecast Accuracy	10%	How well your budgets predicted what you actually spent	78	7.8
Overall	100%	Weighted total, then a letter grade and a GPA	84.9	B

The Grade Scale

Grade	Score Range	GPA
A (A+, A, A−)	90 to 100	3.7 to 4.0
B (B+, B, B−)	80 to 89	2.7 to 3.3
C (C+, C, C−)	70 to 79	1.7 to 2.3
D (D+, D, D−)	60 to 69	0.7 to 1.3
F	0 to 59	0.0

The GPA interpolates within each letter’s band and never crosses into the next grade’s territory, so the letter and the number can never disagree.
Missing data is never graded. A pillar without enough data is excluded and the weights renormalize; the card says “Provisional: 3 of 5.” Early versions scored missing pillars as perfect, so every new user started with an inflated grade that could only fall.
The philosophy fits in one line, stated in onboarding: spend as close to your budget as possible without going over.
Top Fixes ranks the actions that would recover the most weighted points, so the card ends with what to do.

The Rules

Pillar By Pillar

What each pillar counts, what it deliberately ignores, and why.

Each category is scored on how far actual spending sits from its budget.

Over budget is penalized 2 times, under budget 0.6 times at month end, with strict (4 and 1) and relaxed (1.5 and 0.5) presets for Pro users. Larger dollar overages penalize more than small ones.
Mid-month, spending is graded on pace. Fixed costs like rent hit on the 1st and are never prorated; variable categories are paced, with penalties damped early in the month so one grocery run does not look like a blowout.
Savings are one-sided. Saving more than planned never lowers a grade, and moving money into savings is never counted as spending. The failure mode of a naive distance-from-budget score is punishing people for saving too much.
One-time planned expenses can be excused from the discipline pillars, so a planned car repair does not tank a month.

Emergency fund, debt to income and month-end cash, weighted 40, 35 and 25 by default, with Safety Net and Debt Focus presets.

Emergency fund follows the FinHealth steps: 1 month of essential expenses scores 50, 3 months 75, 6 months 100, a straight line between, so the first month counts the most. Cash counts unless it is linked to another goal; investments count at 80% if you turn them on.
Debt to income follows lender practice: 15% or less scores 100, 36% scores 80, 43% scores 60, 50% scores 30, and 60% or more scores 0, with a small bonus when the ratio improves.
Month-end cash: more than 15% of income left scores 100, and ending at zero scores 50.

Households’ spending moves about 30% month to month, so the pillar is built to forgive ordinary variation and catch real spikes.

A spike is a rise above 50% and costs 15 points; steady rises cost 1 point per percent, capped at 25. Spending less than last month never counts against you.
Changes are measured in dollars, not as a share of income. A flat $110 insurance bill used to look volatile whenever income moved.
A floor of $25 or 10% of the budget, whichever is larger, means one charge landing in a different month is not a spike.
A rise that holds the next month is a new normal. A new subscription raises the total once and then stays level. A category with no spending is not graded.

The 50/30/20 framework caps needs and wants and sets a floor for savings. Spending less on needs or wants is never a problem, so the pillar only counts going over.

Saving more than 20% never lowers the grade. Presets shift the split for other situations: 40/20/40 for aggressive savers and 65/20/15 for high cost of living.
Savings under the target still count, so a household that overspends on needs and saves nothing cannot hide behind the other pillars.

The average miss between budget and actual, weighted by each category’s share of the budget, over the last six months.

The first 10% of the average miss is free; past that, the balanced preset costs 1.5 times the excess, with stricter and more relaxed presets for Pro users.
A 20% miss grades a B, since a 20% miss is a good forecast when natural swings are around 30%.

The Audit

Grading Against Published Standards

A grade is only fair if meeting the standard earns an A. I tested each pillar against the benchmarks and found five places the math graded healthy behavior as failing.

Part	What The Standard Says	What The App Did	What It Does Now
Needs vs Wants	50/30/20 caps needs and wants; spending less is never a problem	Under target after month close was penalized like over: saving 60% graded D+, while overspending on needs and saving little graded B	Only over-target counts; saving 60% is A+
Stability	Typical household spending moves about 30% a month (JPMorgan Chase Institute)	A category rising 30% every other month graded F	Spike at 50%, swing penalty halved and capped; a 30% rise earns about a B
Emergency Fund	FinHealth Score: 6+ months is 100, 3 to 5 is 75, 1 to 2 is 50	A straight line to 6 months: 1 month scored 17	The FinHealth steps: 50, 75, 100
Debt To Income	Lenders’ standard is 36% or less for all debt (the 28/36 rule)	A compliant 36% graded F	36% scores 80; 15% or less scores 100
Forecast	Variance practice: 5% fine, 10% review	A 20% miss scored 70 (C−)	10% grace, then 1.5 times; a 20% miss scores 85

38F: a household that overspends every month

84B: a typical household

99A+: a saver

The principle: meeting the published standard earns an A, typical behavior earns a B or C, and only real risk earns a D or F. I kept the school scale because it is the brand, and tuned the parts so standards land at an A.
A calibration test guards the spread. The three households above are fixtures, so no future rule change can quietly inflate or sink every grade.
The benchmark was the Financial Health Network’s FinHealth Score, backed by Vanguard’s emergency-savings research, lender debt-to-income practice and household-spending data. Other budget apps do not publish a comparable grade, so fairness had to come from outside standards.

The Data

Only Honest Inputs Get Graded

A grade computed from inflated numbers is worse than no grade, and real bank data is adversarial.

Sync

Plaid link, sync and webhooks run in server-side functions. No third-party key ever ships in the app.

Classify

Your rules apply automatically. Claude’s suggestions queue for review. Everything else stays uncategorized.

Exclude

Transfers are a third state, outside income, spending and every pillar. Planned one-time spend is excused.

Score

A pure TypeScript module on the device, with one source for every headline figure.

Alert

A server function mirrors the same exclusions, so alerts never contradict the grade.

The transfer problem came from my own bank data. Credit card payments, moves between accounts and card-payment “thank you” credits inflated both income and spending, which synthetic data never would. The fix was structural, a third transaction state, and the same exclusions were then carried into alert evaluation and subscription detection.
AI classifies, the user decides. Only your own rules write without asking. Claude’s results are suggestions with confidence badges and bulk review, and unclassified spend is counted in the grade with a visible “+$X uncategorized” note, because letting it vanish was grade inflation.
Debt payments get their own handling: labeled, linked to their debt, treated as spending, and kept out of the AI pipeline so a loan payment cannot be misfiled as dining out.
Debt payoff tools sit beside the grade, with avalanche and snowball ordering and a “what would an extra $100 do” sheet, so the app can answer what to do as well as how you did.

Money

Money And Trust

Free is genuinely useful. Pro pays for the two things that cost money per user.

$3.99Per month, with the first month free

$39.99Per year

$99.99Lifetime, capped at 500 slots a year

The free tier is the full report card: 21 default categories, savings goals and debt tracking. Pro adds Plaid bank sync, Claude classification, scoring presets, one-time tags, unlimited categories and trends. Bank sync and AI are the two real per-user costs, so conversion comes from them and not from frustration.
Two promotional lifetime paths at $74.99, Founder for early adopters and EDU for .edu addresses, both mapping to the same single Pro entitlement.
The paid flag was client-writable until my own security review found it. Any user with the public key could have marked themselves Pro, and since Pro includes AI classification, a bypass would have burned metered API spend on my key. A database trigger now rejects client writes, and only the signature-verified purchase webhook can set the flag.
Lifetime slots are race-safe. The 500-a-year counter decrements under a Postgres row lock, so two simultaneous purchases cannot both claim slot 500.
Account deletion unwinds the bank connection first. The original code nulled the Plaid token before calling Plaid’s item removal, which would have left the item orphaned and billed forever. Deletion is now a hard delete, verified on a real account.

Retrospective

Why Fixes Kept Coming Back

Each device pass found 40 or more issues, and the fixes added new ones. The way out was finding the causes behind the symptoms.

Cause	Evidence	What Changed
No shared building blocks	24 screens shared 6 components; 45 separate stylesheets; dates formatted 13 different ways; 76 native alert menus in 21 files	A small UI kit and one formatting module, plus a pre-build script that fails on raw font sizes, native alert menus, stray switches and sentence-case labels
Two truths per number	Income read three ways on three screens; 7 of 41 findings in one pass were the same number computed twice	One source per headline figure, a lint rule, and a test that fails if a screen computes its own
Screens showed the engine, not the answer	Raw 0 to 100 scores, “MAPE,” “Shortfall Score”	Status words and a letter grade on every row; the math lives on the website
Fixes were big	203 commits and about 10,000 lines changed in about two days; shared components moved every screen at once	A design freeze until 2.0, severity levels, and approval for any shared-component change
I marked things fixed that were not	A finding reopened after it was called done; some fixes only hold on a device	Fixed means seen on a device in the next build, and the ship bar is two clean passes in a row
Ten styles multiplied everything	Each style in light and dark was 20 sheets per screen	Three styles at launch, swept against a route inventory of 31 routes, 3 styles and 2 schemes
A misconfigured exclude	TypeScript silently skipped checking the whole app for weeks, masking ten real errors	The gate runs the type checker, a migration-grants check and the tests before any commit

Testing against live bank data from week one would have found the transfer state, income promotion and proration logic earlier; they all came from late contact with reality.
Postgres 17 row-level security interacts subtly with soft deletes: a deleted-at predicate in a select policy silently breaks updates on those rows.
Auth flows cost more than any feature. Email confirmation and password reset on mobile produced more beta-blocking bugs than anything else.

Stack

Under The Hood

One Expo codebase for both stores, a Supabase backend that owns every sensitive operation, and third-party services kept strictly server-side.

App

Expo and React Native with TypeScript and Expo Router. The scoring engine runs on the device and is unit-testable in isolation. Live on iOS and Android.

Supabase

Postgres with row-level security on every table, 36 migrations, and 13 Deno Edge Functions: sync, classification, alerts, webhooks, feedback and deletion.

Services

Plaid for bank sync, the Claude API for classification, RevenueCat for purchases and Sentry for crashes.

Expo and Supabase suit one person shipping and operating the product: EAS handles signing and builds, and row-level security replaces a whole backend authorization layer. The trade-off is vendor dependence, with clean exits: it is Postgres underneath, and the purchase model is three products and one entitlement.
Plaid tokens are encrypted and webhook signatures verified, with biometric or PIN app lock and in-app account deletion.
The gate before every commit runs the UI rules check, the migration-grants check, the type checker and 188 tests, backed by on-device flows on both phones.
Marketing site: static HTML on S3, CloudFront and Route 53 at budgetgpa.com.

React Native Expo TypeScript Zustand Supabase Plaid Claude API RevenueCat Sentry Jest Maestro

A score people trust is worth more than a dashboard people ignore.
The through-line with my professional work is exact. The corporate scorecard works because a VP can glance at a C+ and know where to look. Budget G.P.A. bets the same is true for a person and their own money, and that the hard part is not the grading math but making sure the inputs deserve to be graded.

Visit budgetgpa.com

Back to Projects