Live demo | Demo video

Food forecast

Feeding update

Order approval

Replenish is an AI-assisted pet-food agent that turns feeding changes into inventory forecasts, replenishment decisions and budget-controlled PayPal Sandbox payments.

Payments use PayPal Sandbox only. The store, stock, and delivery are simulated, and stock is estimated rather than sensor-measured.

These are feature states, not a completed automatic payment. The video shows a completed Sandbox test transaction. The hosted demo is one shared household, so its state can change.

How it works

From Qizai's two 75g meals per day, send “Feed him 10g less per meal from now on”: the saved result is 65g per meal / 130g per day, with a recalculated forecast and Undo.

flowchart TD
    F[Feeding instruction] --> I[Parse intent and validate]
    I --> P[Deterministic stock and date forecast]
    P --> D{Order due?}
    D -->|No| W[Wait for next check]
    D -->|Yes| G{Stored method, auto authorization<br/>and budget available?}
    G -->|Yes| A[Automatic payment]
    G -->|No| R[Review exact order]
    R -->|Approve and method ready| C[Create and capture payment]
    R -->|Decline or no approval| N[No payment]
    A --> C
    C --> V[Validate result and record status]

Loading

The 8kg / $72.99 USD bag exceeds the default $50/order cap; the monthly limit is $120, including commitments. A fresh, authorized household with an owner-set $100/order cap can take the automatic branch. Approval does not change future limits.

Checks run at startup and every 15 minutes. Food received adds the bag once; delivery is simulated. Follow the repeatable mock walkthrough for both branches.

Quick start

Requires Node.js 22.13+ or 24.x, pnpm 10.33.4 and a browser. Install Corepack if your Node distribution omits it. Use a fresh shell without exported service settings and fresh local state.

git clone https://github.com/HanTirong/replenish-ai-hackathon.git
cd replenish-ai-hackathon
corepack pnpm install --frozen-lockfile
cp .env.example .env
corepack pnpm build:web
node --env-file=.env --import tsx packages/api/src/http.ts

PowerShell: replace cp with Copy-Item. Open http://127.0.0.1:4781; /v1/health returns {"ok":true}. No AI key, PayPal credentials or external database are needed.

The template uses PAYMENT_PROVIDER=mock. pnpm api does not load .env; the explicit Node command does. Exported variables override the file; PORT overrides API_PORT. Leave DATABASE_URL unset.

Mock startup does not establish a saved payment method or auto-pay authorization; the walkthrough supplies simulated prerequisites. Frontend hot reload and the rest of the setup are in local development.

Architecture

flowchart LR
    Web[web<br/>React / Vite] --> API[api<br/>Node.js orchestration]
    API --> Domain[domain<br/>Deterministic forecast and budget]
    Domain <-->|Intent only| AI[DeepSeek<br/>or local rule parser]
    API --> Agent[agent<br/>Payment adapter and journal]
    Agent --> PayPal[PayPal Sandbox<br/>Vault v3 / Orders v2]
    API --> Store[store<br/>Simulated catalog and orders]
    API --> DB[(SQLite)]
    API -.-> Telegram[Optional Telegram notifications]

Loading

Package names correspond to packages/. AI interprets intent; backend rules validate commands and decide purchases. PayPal credentials never enter the LLM. Stock is estimated, not sensor-measured.

Tests and verification

corepack pnpm test
corepack pnpm typecheck
corepack pnpm build:web

These checks use simulated services. Browser verification requires Playwright Chromium; see developer tests.

Documentation

• Local development
• Sandbox setup
• Walkthrough

Security · MIT license · Asset/font licenses