Disclosure: This post mentions my own project, Dial Hanna. I may earn money if shops use it later. The article is a hackathon build report and stands on its own — product links are optional. There are no affiliate / referral links in this post.
On 4 Nov I submitted Dial Hanna to the Meta Global AI Developer Hackathon ($1,000,000 prize pool) — Indie Developer Track, problem statement Build for Your World (team Vibe Coder's Life).
This is what I built, what broke, and what I'd reuse on the next sprint. It is not a sales page.
What Dial Hanna does
Dial Hanna is an AI phone assistant for small businesses:
- You keep your own number.
- When you're busy, call forwarding reaches Hanna.
- She answers from owner-approved knowledge (not an open-ended chat brain inventing shop facts).
- After the call, the owner gets Intent → Outcome → Next action, not a raw transcript dump.
I built it because I live in Budapest and tried to order a cake from a local bakery: email silence, unanswered call, reply two days later — by then I was shopping elsewhere. Same gap hits salons, workshops, guesthouses, trainers, freelancers: hands stay busy, the phone still rings, the job walks next door.
Starting templates I sketched for the submit cover salons, auto & tyre, trades, moving & cleaning, guesthouses, trainers, photographers, pet grooming, hospitality, flowers, language schools, and freelancers — but the agent pattern is “any small shop with a website and a phone,” not a vertical SaaS pitch deck.
Why I keep joining hackathons
Before Meta I joined a Replit buildathon. The format sticks: empty page, you pick the topic, hard deadline. No client brief. Just “what would I ship in ten days?”
For Meta I wanted something a shop could actually benefit from — not a stage demo that dies in a folder. The quiet lesson of the week: with Meta's models, features and apps, a working phone agent is a ten-day build if you keep the scope brutal.
How the product flow works
From the shop owner's side (not the model IDs yet):
- Keep your own number. Turn on call forwarding when you're busy, in a meeting, training, or mid-job.
- Paste your website URL. Muse Wonder drafts a knowledge base — profile, hours, likely FAQs — for the owner to approve. That approved list is Hanna's database. There's a verification call before go-live so forwarding really reaches her.
- When you're unavailable, Hanna answers. She introduces herself as an AI in the first sentence. She takes multi-slot orders and bookings (product, size, pickup time, name, phone), reads them back, and closes politely.
If she doesn't know something, she doesn't guess. She offers a callback, logs a knowledge gap, and the owner's answer becomes an approved FAQ. The knowledge base grows from real calls.
| Step | What happens | Stack |
|---|---|---|
| Onboarding | Website URL → draft knowledge base + FAQs | Muse Wonder |
| On the call: hearing | Caller speech → text (English for now) | Muse Voice Transcribe |
| On the call: answering + booking | Multi-turn orders, notes, Intent/Outcome/Next | Muse Spark |
| On the call: speaking | Hanna's voice | ElevenLabs Flash |
Most AI phone tools start with a blank form. Here the website does the first draft — that was the product bet for a ten-day window.
The Meta stack (and one voice layer)
I wired Dial Hanna on Meta Muse Code / Meta Model API (api.meta.ai/v1). Three Meta models do the agent work; ElevenLabs is only TTS — Meta had no TTS in this path for me.
| Piece | Role | Notes (Nov 2026) |
|---|---|---|
Muse Wonder (muse-wonder-1.0) | Website → knowledge base + draft FAQs | Frontier agentic model for the onboarding draft |
Muse Spark 1.3 (muse-spark-1.3) | Live call brain: tools, orders, notes | Also powered my automated text test suite |
Muse Voice Transcribe (muse-voice-transcribe-1.0) | Speech → text | English only in this build; realtime WebSocket + turn detection; text only |
| ElevenLabs Flash | Spoken voice | One light layer on top of Meta |
| Meta Muse Code | How I built it | Coding agent on Meta Model API |
Limits I'm not papering over:
- STT is English for now. No Hungarian speech-to-text in this build. Spoken voice can still be HU or EN — hearing Hungarian callers is the open gap.
- Muse Spark has to actually call tools for booking slots. If tool-calling flakes, the call flow flakes.
- One Meta API for Wonder, Spark and Transcribe. ElevenLabs only says the words out loud.
The flip side of a ten-day Muse build
I'm not going to pretend the week was smooth:
- Muse Code in the terminal can be very slow on thick specs and bigger changes. When the prompt got dense, I waited a lot.
- Demo video render + transcript took hours. Everything else was ready during the day on 4 Nov; I submitted around 10–11 PM Budapest, a couple of hours before the 9:00 PM ET deadline.
- Long-context loss on Muse Spark 1.3. In a long session it lost prior conversation, specs and markdown and drifted. On the hackathon feedback form I logged the biggest obstacle as the model not following instructions.
None of that kills the thesis — Meta's stack still got a working phone agent shipped in ten days — but it's the honest flip side if you're planning a similar sprint.
How I'd pick models after this week
Verified list prices as of 6 Nov 2026:
| Model | Input / output (per 1M tokens) | Context |
|---|---|---|
| Muse Wonder | $4 / $20 | 1M |
| Claude Opus 5.5 | $4 / $20 | 1M |
| Gemini 4 Argon | $2 / $10 intro through 30 Nov 2026; then $4 / $20 | 1M |
Sources: Meta Model API pricing, Anthropic pricing, Gemini API pricing.
My take (builder experience, not a ranking law):
- Claude Opus 5.5 currently leads for me on coding-agent autonomy and terminal use. I did not see the same context-loss / slow-reply pattern with Opus 5.5 that I hit with Muse Spark 1.3 in a long Muse Code session.
- Gemini 4 Argon currently leads for me on multimodal / document retrieval jobs.
- Muse Wonder is mainly useful when you're inside Meta workflows and Meta products (Facebook, Instagram, WhatsApp) — which is exactly why it fit a Meta hackathon phone agent.
For Dial Hanna I stayed on Muse because the brief was Meta. For day-to-day coding agents outside that lane, Opus 5.5 is where I'd start right now.
EU builders: disclose the AI
Since 2 Aug 2026, Article 50(1) of the EU AI Act requires AI systems that interact directly with people to tell them they're talking to an AI, unless that's obvious. The European Commission says people must be told from the start of the first interaction. On Dial Hanna, Hanna introduces herself as an AI assistant in the first sentence of every call. If you're building a phone agent for the EU, put the disclosure in the greeting — not in a FAQ after the fact. (European Commission FAQ)
What I'd tell you before your first hackathon
- Pick a problem you watched happen. Mine was a bakery that couldn't answer while I waited on a cake.
- Keep the plan tight. You still have to deliver a finished product by the deadline — scope is the whole game.
- Budget time for demos and renders. Video + transcript burned hours I did not leave enough buffer for.
- Watch how the deadline is structured. Some buildathons stage plan / build / publish. Meta's Global AI Developer Hackathon this time was a one-off submission deadline.
- Lean on the models you're given. Wonder for the website draft, Spark for the call, Voice Transcribe for hearing, Muse Code to wire it — that split made the sprint possible.
Status (short)
Submitted 4 Nov, Indie · Build for Your World · team Vibe Coder's Life. Judging roughly 4–8 Nov, winners around 9 Nov. Hungary-first build; English STT for now. Personal next milestone: first real customer — separate from the hackathon score.
If you want the longer version with the embedded demo video and FAQ, it's on my site: vibecoderslife.com/post/dialhanna-ai-assistant-meta-hackathon. Optional context from the apply post: I got into Meta's $1M AI hackathon.
