DealRoom

Talk, agree, pay with PayPal and deliver in one chat, where every step is on the record. If you did nothing wrong, the record proves it.

Comment

Inspiration

Freelancers work across a dozen apps: the client on WhatsApp, the brief in email, the file on Drive, the money on PayPal, Wise or a bank transfer. When something goes wrong ("I never received it", "that's not what we agreed", "pay me on Friends & Family to skip the fees"), it's their word against the client's, and the proof is scattered everywhere. I wanted one place where the deal itself is the proof.

What it does

DealRoom is a chat where people who sell digital work and people who buy it talk, agree, pay and deliver. Every step is recorded with its time.

  • The chat becomes a contract. When the conversation has the work and a price, the AI drafts the terms from the chat. Both sides accept, and only an agreed change can move them later. New work becomes a new deal.
  • PayPal payment locked to the terms. The freelancer makes a real PayPal invoice, and the client pays with the PayPal button. The amount comes from the accepted terms, not from the browser.
  • AI file check. Every delivery (image or PDF) is checked against the terms: "Translated to English ✗", "Yellow background ✗".
  • Download record. When the client opens the file, it's recorded, which matters when someone says "I never received it".
  • Scam warnings. Off-platform payment tricks (Friends & Family, Wise, crypto, "I overpaid, send the difference back") are flagged as they're typed.
  • Private honesty check. Before someone sends a claim the record contradicts, DealRoom shows only them the facts, so nobody gets accused by mistake.
  • Grounded dispute evidence. One click builds a report for PayPal: a neutral timeline, both sides' statements and an assessment. Every sentence must cite a recorded event, or code removes it. For the freelancer it goes through PayPal's Disputes API; the client gets the PDF for their case.
  • A business dashboard on AG Grid's AG Studio. What you earned through PayPal, who owes you, what needs you now, and "Proof on record" for every deal: terms, payment, delivery, download. An AI assistant answers questions from the record, like "What's my client return rate?"

How I built it

  • Backend: Python, Flask and SQLite. Each visitor gets a private copy of the demo world, and both logins share it, so two windows update live.
  • PayPal (sandbox): Orders v2 with a server-set amount, capture, Invoicing v2, Refunds, the Disputes API to provide evidence, and verified webhooks.
  • AI: NVIDIA Nemotron models on Nebius: Nano for quick classification, Super for terms, claims and dispute reports, and Super VL for checking images.
  • Rule: the AI reads, code checks. Every number in the terms must appear in the chat. Every sentence in a dispute report must cite an event. Every message from the AI-played people is checked against the record before anyone sees it.
  • Dashboard: AG Studio (AG Grid) with custom widgets, a custom theme, and its AI agents connected to the same models. The assistant gets exact facts computed by code from the record.
  • Testing: 61 automated tests, scripted tricky scenarios on the real models (35 of 35 passing), and an AI tester: a model plays a liar, a scammer, a haggler or a prompt-injection attacker against the app, and a second model audits the result.

Challenges I ran into

  • AI that agrees too easily. The AI-played client said "I've sent the payment" when PayPal showed nothing, and offered to pay a deposit twice. Now code checks every AI message against the record, and the model retries or stays silent.
  • A curly apostrophe broke a safety check. "I’ve" slipped past a rule written for "I've". The AI tester found it.
  • Terms that followed the first request instead of the latest agreement. "I don't do brown, only red" still produced brown terms until the rules put the latest agreement first.
  • A vision model fooled by a detail. A blue dog portrait passed "red logo" because the dog wore a red collar. The check now confirms the file is the agreed item, and that a color means the main colors of the work.
  • A payment that silently didn't complete. PayPal refuses to capture an order whose invoice id it has already seen. Every order now gets a unique id, and a failed capture is shown, never swallowed.

Accomplishments that I'm proud of

  • Real PayPal payments, invoices and dispute evidence, all tied to one recorded conversation.
  • AI that is checked by code instead of trusted: no invented payments, no contracts that drift from the chat.
  • Letting an AI attack my own app, and fixing what it found.

What I learned

Trust in a marketplace isn't about being nice. It's about proof that both sides can see. And with AI, the useful question isn't "can it write this?" but "how do I check what it wrote?"

What's next for DealRoom

  • File previews that unlock fully on the final payment, so the freelancer never gives work away for free.
  • Payment reminders sent through PayPal Invoicing.
  • Live updates in the dashboard the moment a PayPal payment lands.
  • Real accounts and invitations, so any freelancer can bring their clients.

Built With

Share this project:

Updates

Submission history