Skip to content

DEVPOST.md

Latest commit

 

History

History
97 lines (82 loc) · 4.61 KB

DEVPOST.md

File metadata and controls

97 lines (82 loc) · 4.61 KB
Raw

Zeroth — Devpost submission

Elevator pitch: Zeroth locks your money to its purpose and lets an AI agent spend it in plain language — it can pay your bills, but by design it can never cash out.

Try it out

About the project

Inspiration

The average Nigerian doesn't struggle to earn small money — they struggle to keep it. Money meant for rent or data quietly leaks into impulse spends, and the only "savings locks" available are blunt: either you can't touch the money at all, or you can withdraw it the moment you're tempted. There was no middle ground where money stays useful (you can still spend it on what it's for) but can't be impulsively cashed out.

The name comes from the Zeroth Law of Thermodynamics: if two bodies are each in equilibrium with a third body, they're in equilibrium with each other. We put an AI agent as that third body between you and your money — every transaction passes through it, so your rules always hold.

What it does

You lock an amount you commit not to withdraw, and split it across purpose wallets (Savings, Rent, Bills, Food, Airtime, Data — or your own custom ones). Then you just talk to your agent:

  • "Buy ₦500 airtime" → it tops up your line.
  • "Abeg buy me 1k data" → it understands Pidgin.
  • "Send ₦5k to 08145997956 PalmPay Samuel" → it asks "what's it for?", then executes from the right wallet once you say "for feeding".
  • "Cash out ₦40k to my bank" → 🔒 blocked, by design — locked money can't leak.

How we built it

  • Frontend: Next.js 14 (App Router), TypeScript, Tailwind, shadcn/ui, Phosphor icons, a monochrome design system with smart light/dark theming.
  • AI: Google Gemini (gemini-flash-latest) does two jobs, both with strict structured (JSON-schema) output:
    1. Intent parsing — turns messy natural language (Pidgin, slang, "5k", worded amounts) into a structured PaymentIntent.
    2. AI money split — from your stated priorities ("school fees, business, black tax") it proposes purpose wallets that sum exactly to your total.
  • The key idea: the AI only ever produces language → structure. Every decision about money — which wallet funds it, per-transaction limits, balance checks, and the absolute refusal to cash out — is made by a deterministic policy engine. That's what makes it safe to let an AI move money: the model never decides anything about the money.
  • Multi-turn agent: transfers with no stated purpose are held as a pending intent while the agent asks "what's it for?", then completed on your reply.
  • Deploy: Vercel, self-contained demo mode (no login) so anyone can try it instantly; Gemini has an offline fallback so the demo survives bad wifi.

What we learned

  • The cleanest way to make an LLM trustworthy with money is to give it no authority over money at all — separate "understanding language" (AI) from "moving money" (deterministic rules).
  • Gemini's structured-output + schema makes it reliable enough to sit in a real product loop, not just a chat box.
  • Multi-turn state (ask → remember → complete) turns a rigid command parser into something that feels like an assistant.

Challenges we faced

  • Distinguishing a peer transfer from a cash-out — "send ₦5k to a person for feeding" must be allowed (spend), while "withdraw to my bank" must be blocked. We solved it by requiring a purpose to pick a funding wallet and asking for it when missing.
  • Making the AI split always sum to the total — the model returns percentages; we normalise and fix rounding so wallets add up exactly.
  • Flaky venue wifi + Gemini quotas — we added an offline regex fallback and pinned the working model, so the demo can't die on stage.
  • A cohesive monochrome redesign — light + dark themes with no flash, driven entirely by design tokens.

What's next

  • Real payment rails (VTpass / Reloadly / Flutterwave Bills) behind the same interface.
  • Persistent multi-user accounts and per-category autonomy budgets.
  • Bank-statement onboarding to infer your starting split automatically.

Built with (tags)

next.js · react · typescript · tailwind-css · shadcn-ui · google-gemini · gemini-api · generative-ai · ai-agent · fintech · programmable-money · policy-engine · vercel · web-speech-api · phosphor-icons · node.js · nigeria · pidgin · natural-language-processing · structured-output

Team

  • Inioluwa Aroyewun — Full-stack & AI engineering
  • Nuffi Akolade — Product, design & pitch