Inspiration
Small studios and their clients rarely fight about the amount. They fight about whether the thing was actually done. "You approved it." "We approved a preview." An invoice cannot settle that, because by the time there is an invoice the reason for the payment has been lost. We wanted the reason to travel with the money.
What it does
- New: Talk it through: record or paste the conversation where you agreed, and it becomes a one-page pact where every term opens the sentence it came from. Fix a term by saying it ("make it two payments"). Milestones are released against a photo of the finished work. PACT turns a plain-language agreement into commitments that each cite the exact words they come from.
- Select a sentence in the agreement and record a commitment against it, with an amount, a reviewer and a completion criterion.
- An optional AI step answers one kind of question: "which words cover this?" It returns an exact passage from the agreement, quoted by character position. It never sets an amount, accepts evidence or decides anything.
- The performer logs delivery evidence, including the actual file. The reviewer can check their own copy against the recorded one.
- Before the reviewer accepts, PACT shows what that decision will unlock, including dependent commitments.
- Only an accepted commitment can become a payment. PACT creates a PayPal Sandbox order for exactly that amount, the payer approves on PayPal, and a separate, deliberate click captures it.
- Disputes, revocations and deadlines are separate dated events, and readiness is always recomputed from the history.
In the demo, a $425.50 commitment is blocked because only a preview was delivered. The final WAV masters are logged, the client's reviewer accepts, and a real Sandbox order is approved with a test card and captured.
How we built it
Node.js with no runtime dependencies and a browser app in plain HTML, CSS and JavaScript modules. PayPal REST in Sandbox: OAuth and Orders v2 create, read and capture, with a server-side journal so an uncertain provider result is never reported as success or blindly replayed. The AI step calls any OpenAI-compatible chat endpoint; the demo uses a local Qwen 2.5 7B model through Ollama. Files are fingerprinted in the browser and never uploaded.
Challenges we ran into
Making a model useful in a setting where a wrong word matters. We restricted it to pointing at text that already exists in the agreement and reject anything that is not a literal passage. Handling the moment after a payment request when the answer from the provider is unknown also took care: PACT keeps "unknown" as a real state.
Accomplishments that we're proud of
A complete path that holds together on real Sandbox payments: sentence, commitment, evidence, decision, order, capture. The acceptance review that shows consequences before the reviewer commits. An AI feature that is helpful and cannot overreach.
What we learned
People trust a payment more when they can see why it exists. The best use of AI here was narrow: finding the right sentence fast, and leaving every decision to a named person.
What's next
Invited reviewer links so the client acts in their own browser, partial acceptance and amendments, and exports a mediator or accountant can read without the app.
Built With
- css
- html
- javascript
- node.js
- ollama
- openai-compatible-api
- paypal
- paypal-orders-api
- paypal-sandbox
- qwen
