Freelancers and small sellers lose time to invoicing: open a dashboard, fill in a form, copy an amount, paste an email, send. The PayPal AI Invoice Assistant removes the form entirely. You write what you want in plain language — "Invoice Budi for $120 for logo design work" — and the app does the rest:
- A local LLM (Ollama, qwen2.5:3b) extracts the amount, currency, recipient and description as structured JSON.
- The app creates a real PayPal invoice through the PayPal REST API (POST /v2/invoicing/invoices, sandbox) and returns the invoice ID, status and payment link.
- If the local model is unavailable, a deterministic parser takes over automatically, so the app never blocks on AI latency.
Everything runs locally with zero paid dependencies: Python stdlib only, Ollama free and local, PayPal sandbox free. No data leaves the machine except the PayPal API call itself.
Demo mode: without PayPal credentials the app still runs end-to-end and returns a clearly-labelled MOCK invoice (yellow "DEMO MODE" badge in the UI, INV2-MOCK-… IDs, explicit note in the API response). It is never presented as a real invoice. Set PAYPAL_CLIENT_ID / PAYPAL_CLIENT_SECRET to switch to the live sandbox.
How to run
ollama serve & ollama pull qwen2.5:3b # local AI (optional; fallback works without it)
python3 server.py # → http://127.0.0.1:8080
# Live sandbox:
PAYPAL_CLIENT_ID=... PAYPAL_CLIENT_SECRET=... python3 server.py
Open http://127.0.0.1:8080, type a request, click Create invoice.
AI tools used
- Ollama + qwen2.5:3b (local LLM) — structured extraction of payment details from natural-language requests (JSON-only prompt, temperature 0).
- Deterministic regex parser (built-in) — fallback when the local model is slow or unavailable; every parse response reports which parser ran.
- Muse (Meta AI) — assisted in writing the code and documentation.
