Branches

What inspired us

Fidelity's 2026 Couples & Money study found that 69% of couples aren't regularly talking about long-term finances, even though 41% wish they were. Nearly half avoid money conversations to prevent arguments.

The couples apps we looked at are good at showing where the money went. We didn't find one that answers the question a couple actually asks when life changes,"We're having a baby. Now what happens to the house?"

What it does

Both partners join a household. Branches reads their transactions and works out income, bills and loans, so the partner only has to tap to confirm. They add goals to a shared ten-year plan.

When something happens, they tell Branches. If a goal no longer fits, both partners are pulled into a live session with three paths forward. They can see each other's picks, react, adjust a path together, and pin what matters most if they disagree. Nothing becomes the plan until both press and hold at the same time.

Each partner's fun money stays private, the other sees the amount, but not the purchases.

How we built it

App: one Expo (React Native) codebase for phone and web
Backend: SpacetimeDB, which keeps both screens in sync and runs all the plan logic inside its reducers
Math: a deterministic TypeScript engine
Assistant: Google Gemini, which interprets, names and explains, never does any of the calculations on its own

Every number on screen comes from the engine, and you can tap it to see how it was worked out.

Any AI reply containing a number the engine didn't produce is thrown out, and replies can't contain advice language. Branches shows projections, never recommendations.

Challenges we ran into

One cent broke our first test. Floating point turned $5273.65 into $5273.64, so we now normalize precision before rounding.
The product was hiding behind a button. A session only opened if one partner tapped at the right moment, so we moved the whole chain to the server.
A valid API key looked invalid. An old key in a shell profile was silently overriding our .env.
We changed course mid-weekend. We switched AI providers, gave up a sponsor track to do it, and dropped an iMessage version.
Real phones were hard to test. A carrier hotspot blocked phone-to-laptop traffic.

What we learned

Keeping AI to words made the product easier to trust and easier to test. A model mistake can't change anyone's money.
Writing specs and pass/fail checks first made AI-assisted development far more reliable. The tests caught problems a demo never would have.
The hard part of a couples app isn't the math. It's making sure both people are in the room when the decision happens.