Inspiration

Only 63% of US adults can cover a \$400 surprise without borrowing. Americans paid over \$12B in overdraft fees in 2025, and 47% of pay-later users paid late last year.

The problem isn't just how much money people have. It's that their money is scattered across checking, savings, cards, loans and pay-later plans, and every banking app shows the past. Nobody tells you that rent on the 1st will take you below \$0 until it already has.

We built Ahead for people like Maya: 24, first job, \$1,100 rent, a card that's 59% full, and an overdraft coming she can't see yet.

Banks show your past. Ahead shows what's next.

What it does

Ahead has three pillars: See, Predict, Act.

  • See: checking, savings, card, auto loan, pay-later, 401(k) and Roth IRA in one place, plus a health score and a credit score.
  • Predict: a 60-day forecast with a safe-until date, so you know the exact day money gets tight.
  • Act: every alert comes with three fixes and a before/after preview. One tap (or a spoken "yes") makes a real transfer through Capital One's Nessie API, and every action can be undone.

You can also ask anything by text or voice ("Can I afford a trip in March?"), upload a bill or pay-later screenshot, cancel unused subscriptions, and set up autopilot savings.

Nothing moves without your OK.

How we built it

Frontend: Next.js, TypeScript, Tailwind, framer-motion. Backend: FastAPI (Python) with pandas and numpy.

Layer Technology
Planner NVIDIA Nemotron nemotron-3.5-super-vl-120b-a12b
Voice NVIDIA Parakeet (speech to text) and Magpie (text to speech)
Documents Google Gemini 3.5 Flash
Bank Capital One Nessie v2 (real reads and writes)
Ledger Tiger Data (Postgres + TimescaleDB)

Python does the math, AI only plans

Our core rule: the LLM never invents a number. Nemotron decides which tools to call, and Python does all the math. A critic then checks that every number in the answer matches a tool output. If anything doesn't match, the answer is rejected.

The forecast

We run a Monte Carlo simulation with $N = 1000$ paths over 60 days. Each day's spending is resampled from Maya's real history, and bills and paychecks land on their exact dates:

$$ B_{t} = B_{0} + \sum_{d=1}^{t} \left( S_d + F_d \right), \qquad S_d \sim \text{bootstrap(past daily spend)} $$

The safe-until date is the first day the median balance drops below a \$100 buffer:

$$ t^{*} = \min {\, t : \operatorname{median}(B_t) < 100 \,} $$

The uncertainty band is the 10th to 90th percentile of the 1,000 paths.

The health score

Four indicators, each worth 0 to 25 points and scaled linearly:

$$ \text{Score} = \sum_{i=1}^{4} 25 \cdot \operatorname{clip}!\left( \frac{x_i - z_i}{f_i - z_i},\, 0,\, 1 \right) $$

The four indicators are the safety cushion, pay going to debt, savings rate and how full your card is.

Daily spend limit

$$ L = \frac{\text{checking} - \text{bills before payday} - \$100}{\text{days until payday}} $$

Challenges we ran into

  • The LLM was slow and sometimes wrong. Answers took 10–36 seconds, and some came back blank but still "verified." We added a fast path that answers common questions with exact math in under 1.5s, turned off long reasoning for Nemotron (about 7s now), and made blank answers fail the check.
  • Nessie v2 can't backdate transactions. So we split the data: Nessie holds the live accounts and every approved move, and Tiger Data holds the 90-day history.
  • Free-tier limits. Gemini allows 20 requests a day per model, so we added a backup model and stopped retries from wasting quota.
  • Making the numbers believable. We checked Maya's salary against her take-home pay, and her loan against its monthly payment, so every figure holds up in real life.
  • Keeping the demo unbreakable. Every outside service has a timeout, retries and a fallback. With no keys at all, Ahead still runs on built-in data.

What we learned

  • Trust beats cleverness. The Verified badge and the "How I figured this out" trace mattered more than flashy AI answers.
  • Let each part do what it's good at. LLMs plan and explain well, and Python does math reliably.
  • Warn early, fix fast. The best money app speaks up before the problem, with a fix ready.
  • Simple wins. One number (safe until), one next step, plain words, no shame.

What's next

  • Spanish with Nemotron NMT
  • Unusual-charge and fee alerts
  • Real virtual-card controls for cancelling subscriptions

Ahead keeps you one step ahead of your money.

Built With

Share this project:

Updates

Submission history