# Vaikhari AI

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Satyam Dwivedi](https://cerebralvalley.ai/u/vaikhari-ai)
- **GitHub:** https://github.com/vaikhariai/delegate
- **Demo video:** https://youtu.be/JH_sPynsC34
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2066

Delegate lets an AI agent negotiate a marketplace sale for you, inside limits the server enforces. Live demo (fictional sandbox, no sign-in): https://delegate-ybzssouqyq-uc.a.run.app

Why: in September 2026 a Toronto seller let Meta's Muse agent handle a Marketplace sale. It shared his home address through a reply template he had set to "always allow", sent "Yep, I'm here!" from his account when the buyer arrived, and on another listing appeared to offer $100 under his minimum; Meta acknowledged the pricing error to AP. The model was capable; nothing outside it enforced what the seller had agreed to.

How: the seller sets a mandate (asking price, private minimum, pickup only, cash, which facts may be shared) by form or by a voice note in English or Hindi, drafted by Muse Spark 1.2. For every buyer message, including Hinglish, Muse Spark 1.3 proposes one typed action through structured output. The server checks it in ten fixed steps and sends only fixed template text, so the model never sees the minimum and cannot promise anything outside the mandate. A reservation needs the buyer's explicit Accept and the seller's approval of that exact action in the seller console, never a "yes" in chat; the server writes it once, reads it back, and only then marks it Verified. The seller can revoke at any moment.

Results (the claim is safety, not accuracy): in 106 live negotiations with Muse Spark, Delegate made 0 forbidden commitments. The same model with the mandate checks switched off, in a private evaluation only, reserved for a scam buyer 8 times and, having no approval step, 2 times without the seller's approval. On the harder language probes a keyword parser with the same checks passed 2 of 24; Muse Spark passed 23 of 24. The gate matches 151/151 cases written from the policy, and 290 automated tests pass. A small pilot, not a statistical estimate.

Who pays: sellers who list many items, and platforms' trust and safety teams that need auditable agent controls. Limits: the buyer chat is simulated (a WhatsApp Business adapter is implemented and fixture-tested, not connected to a live number), and everything runs in a fictional sandbox. Built during the hackathon; open source under Apache-2.0.

## More from Meta Global AI Developer Hackathon

- [Marstians](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2063)
- [Ziba Intelligence](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2064)
- [Yuvraj Sharma](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2065)
- [Practice Sheets](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2067)
- [MOLLI](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2068)
- [Fabulka](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2069)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/2066. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt