a1mobile Voice AI 1 day Hackathon - $4000 in Prizes
Close the Loop — A Voice AI Hackathon
Build a voice agent that gets a real task done — and prove it.
12 hours · Teams of 1–3 · AI-assisted coding fully allowed · Powered by a1mobile
The one-line version
Anyone can wire up speech-to-text → LLM → text-to-speech in an hour. That's not the game. The game is the action layer: a voice agent that operates against systems and people it doesn't control, completes a real task, and stays honest when reality pushes back.
The loop only closes when we can independently see the result — the SMS that actually landed, the booking that actually hit the system. Not what the agent says happened. What did.
What you'll build
A voice agent that completes a real-world task involving a party it doesn't control — a human, a phone tree, or a live system — and produces a verifiable side effect we can confirm on our end.
Inbound (your agent answers) or outbound (your agent calls out) — your choice.
Pick a task, or pitch your own:
📅 Book / reserve / schedule — a table, an appointment, a service slot — and confirm it
🔁 Change / cancel / dispute — modify an order, cancel a subscription, reschedule
🎯 Get-it-done errand — an underspecified goal across 2–3 tools ("get 4 people a dinner reservation downtown around 7 and text the group the details")
💬 Extract / negotiate — call to get a quote, hours, or availability, and report it back accurately
But here's the twist: every agent faces the same planted friction during judging. The first option is unavailable. A payment fails. The request is ambiguous. The person on the other end interrupts and mumbles. Your agent has to adapt, offer alternatives, and stay honest — a confident lie loses to an honest "I couldn't do that."
Beat the baseline ⏱️
At kickoff, we'll onboard a brand-new business onto a1mobile live on stage, in about 5 minutes — real phone number, working AI assistant, real SMS confirmations. Then that stock assistant faces the exact scenario you will, and its score goes on the board.
It took us 5 minutes. You have 12 hours. Beat it.
Three ways to build
On a1mobile rails (recommended, fastest) — claim a ready-to-go phone number + credentials at kickoff, point it at your stack, and use our tools to send the real confirmation SMS. Skip the telephony setup pain entirely.
Bring your own everything — any provider, any framework. Fully open.
Mix and match — whatever ships.
We provide a starter kit ("hello, you're on a real call"), a sandbox booking system, the difficult-counterparty phone lines, and a live dashboard where your results show up.
🏆 Prizes
Top 3 teams (cash + months of a1mobile service — unlimited calls & texts — + AI credits):
🥇 1st — $2,500 + 3 months a1mobile + AI credits
🥈 2nd — $1,000 + 2 months a1mobile + AI credits
🥉 3rd — $500 + 1 month a1mobile + AI credits
Plus category awards:
Best on a1mobile — deepest, most effective build on the a1mobile rails
Best on VoiceOS — purpose-built use of the VoiceOS MCP: your agent closes the loop into desktop apps (a calendar event, a Slack message, a task filed), not just "an MCP was called"
Most Human — the agent that felt the most natural to talk to
🎬 Video Bonus — post a reel/short about building with VoiceOS (a1mobile welcome too) within 48h after the event, for an extra prize
How you're scored
Judged on a live call we run against your agent — a fresh scenario with the same planted friction for everyone. We score the side effect we can confirm, not the agent's claims.
One hard rule: a fabricated success is an automatic critical flag. If your agent says "all booked!" and nothing happened on our end, it can't place — no matter how good it sounded. Honesty about failure beats a confident lie, every time.
Schedule — Friday, July 31 (9am–9pm)
What to bring
A laptop and your team (1–3 people — solo is fine)
Your own model API keys, or use the sponsor pool (heads up: streaming speech burns credits fast — budget for it)
Curiosity about voice agents that actually do things
No prior voice-AI experience required. The starter kit gets you to a live call fast, so you spend your time on the interesting part.
The rules that keep it fair & legal
AI tooling is unlimited — copilots, agents, any model, any framework. Assumed, not cheating.
Controlled endpoints only. Every counterparty is provided by us — a number we answer or a sandbox we run. No cold-calling real businesses or people. No real financial transactions. This keeps it legal, ethical, and fair — everyone faces the identical planted friction.
The side effect must be real. "Pretend it sent" doesn't count.
No pre-recorded or Wizard-of-Oz demos. Judging is live, against a scenario you haven't seen.
Original work during the event. Starter kits and open-source frameworks are fine; pre-built task-specific agents are not.
Recording notice
Judging calls are recorded and transcribed for scoring, tie-breaks, and event highlights. Registering and attending means you consent to this — full terms at registration. See the Terms & Conditions you'll agree to when you sign up.
Questions? Reach out to the organizers. Rename "Close the Loop" to taste. See you there. 🎙️