About Wallet‑Voice Guardian
Inspiration
Paying bills is usually fragmented: notifications in one place, payment apps in another, and little clarity about what is urgent or safe to approve. We wanted to explore a better interaction model: what if a person could simply ask, “Which payment needs attention?” and receive a clear, grounded answer—without giving an AI the authority to move money on its own?
Wallet‑Voice Guardian is our answer: a voice-first AI payment copilot that helps users understand requests, prepare an exact payment review, and keep approval firmly in human hands.
What it does
Wallet‑Voice Guardian brings pending payments, deadlines, wallet activity, and payment controls into one interface. A user can ask about urgent payments through voice or text, follow up with context such as “pay that one,” and switch the intended funding source to PayPal Sandbox when appropriate.
Before anything is executed, the Payment Guardian review explains:
- Merchant, amount, currency, and deadline
- Payment source and verification state
- Applicable wallet policy checks
- Why approval is required
- Current processing and receipt status
For internal demo-wallet payments, the app uses a transaction-bound wallet PIN, reservations, idempotency controls, policy evaluation, and a double-entry journal. For PayPal Sandbox payments, checkout remains separate from internal simulated funds: the buyer must approve through PayPal, and completion is only shown when provider-confirmed capture evidence is recorded.
How we built it
We built Wallet‑Voice Guardian with:
- React, Vite, and TypeScript for the responsive frontend
- FastAPI, SQLAlchemy, and PostgreSQL for the backend
- A double-entry journal and reservation-based payment execution model
- NVIDIA-ready voice integration boundaries for transcription, reasoning, and speech
- PayPal Orders v2 Sandbox integration for external checkout preparation and capture handling
- Playwright browser journeys and PostgreSQL-backed tests for payment safety scenarios
The assistant retrieves real backend payment data rather than inventing merchant details or balances. It can interpret intent and prepare actions, but it cannot bypass PIN approval, PayPal buyer approval, policy rules, or provider evidence.
Challenges we ran into
The hardest challenge was balancing a natural conversational experience with financial safety. “Pay that one” is easy for a person to understand, but the system must resolve the correct request, handle multiple matches, reject expired requests, avoid currency assumptions, and never turn ambiguity into a payment.
We also had to clearly separate simulated internal-wallet funds from PayPal Sandbox funding. A PayPal capture must not alter the internal wallet ledger, and an OAuth connection alone is not proof of a completed payment.
Another challenge was making complex safeguards understandable. Instead of exposing users to technical terms like idempotency and reservation locks, the interface explains the outcome in plain language: what was checked, why a payment can proceed or is held, and what approval is still needed.
What we learned
We learned that trustworthy financial AI is not about making payments fully autonomous. It is about reducing cognitive load while preserving user control.
The best assistant experience is grounded in real data, explicit about uncertainty, careful with sensitive information, and honest about what it can and cannot verify. Wallet‑Voice Guardian demonstrates that voice can make payments feel simpler without making them less secure.
Built With
- agents
- ai
- docker
- fastapi
- idempotency
- ledger
- nvidia
- paypal
- playwright
- postgresql
- python
- react
- sandbox
- security
- speech-to-text
- sqlalchemy
- text-to-speech
- typescript
- vite
- voice
