# Rabin Singh

- **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:** [Rabin Singh](https://cerebralvalley.ai/u/Rabins)
- **GitHub:** https://github.com/rabinsingh123/dockhand
- **Demo video:** https://youtu.be/GgCGXzVEkE8
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5109

Partners deliver more than 70% of the world's IT spending, US$3.4 trillion a year (Canalys, 2023). Those deals are agreed in chats; WhatsApp alone carries 7 billion voice messages a day. The paperwork that turns a chat into money lives in the big company's portal, and the rules there reward whoever files first and complete: Dell's terms say the partner must be the first to submit the deal, and since November 2 Microsoft gives a partner five days to answer a referral before it can go to someone else. Big companies have integration teams for this. Their partners and suppliers have a login, so the person who closed the deal retypes it at night, and the bot the firm once paid for died the first time the portal changed. I lead GTM engineering and watch this every week.

Dockhand lets the rep forward the message instead. You show it a task once, as a screen recording or just a goal. Muse computer use does it the first time and records exactly what it touched. That run compiles into a program that replays in seconds with no model calls. Before every submit, Dockhand reads every field back from the page and waits for the rep's Approve tap in WhatsApp.

When a portal changes, Dockhand repairs the broken step with the cheapest fix that works: another stored selector or fingerprint similarity (both free), then Muse Spark Open choosing the element from a text-only list of the page, then Muse computer use with at most 8 actions. When the page needs information the program doesn't have, such as a new required field, it asks the rep once and remembers the answer. A nightly watch runs the same repairs on dry runs. At 3:13 AM on Nov 3, after a scheduled change to our partner portal, it sent this on its own: "Caravela Partner Hub changed overnight. I repaired 4 steps (1 free, 3 with Muse Spark Open, $0.0058) and checked it with 5 dry runs. v4 is active, v3 is kept. Nothing to do."

The idea in one line: agents re-think every click on every run, and recorded bots die when the page changes. Dockhand learns a job once, turns it into software, and keeps that software alive.

## More from Meta Global AI Developer Hackathon

- [Hygge](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5106)
- [AURA RLVE](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5107)
- [Slate HQ Ltd](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5108)
- [Ive](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5110)
- [Mantis Analytics](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5111)
- [BSDEV](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5112)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/5109. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt