Grounded voice AI for organisations with a duty of care to the people who call them.

A phone call works where data does not.

 

When someone calls their housing provider about a repair, a helpline about their case, or an event about getting home safely, the answer has to be correct, grounded, and auditable — not a confident guess. These are publics an organisation has a duty of care toward: many callers, many languages and accents, often stressed or in degraded conditions, with stakes too high for a hallucinated answer. A generic voice assistant that makes things up is worse than no line at all. Today that cost hides inside signage, staff, and the info-desk queue.

Dial a local number. Ask a question. Get a grounded answer, in the languages your callers actually speak. No app, no signal needed.

The phone line reaches people that other channels fail to reach, and it is the channel that survives when data networks saturate. Booth answers from an approved knowledge base and is built not to guess — every call returns outcomes: a correctly answered caller, structured intelligence for the organisation, and a clean handoff when a human is needed.

Deliverables are correctly answered calls — grounded, not guessed — with escalations that happen only when a human is genuinely needed and are handed off cleanly, plus the structured intelligence and metrics behind them. Each customer's data stays in its own store, never pooled or benchmarked across customers.

Priced per line and outcome, never per minute. Free for the people who call.

Staying intelligible and grounded over a phone line under adverse conditions — strong accents, background noise, stressed callers, narrowband audio — is the real engineering: turn detection, source-bound answers, robustness, and uptime under load. This is an infrastructure problem, not a prompt wrapper.

• Live telephony. A real inbound number per deployment, over carrier PSTN/SIP. The full loop runs on phone infrastructure, not a browser demo.
• Voice plane. Real-time call handling and agent orchestration run on a self-hosted LiveKit stack on our own EU infrastructure (DigitalOcean, Amsterdam). Deepgram for speech-to-text and Cartesia for text-to-speech; the voice layer is swappable in code, so Booth is not tied to any single provider.
• Cloudflare-first control plane. The backend API, per-customer databases, and web apps run Cloudflare-first and EU-pinned, with each customer isolated in its own store and no raw call audio recorded by default.

Built with  

• A working dial-in prototype answers today on a real number, first built in under 24 hours at EurHackNL.
• Landing site live at dialbooth.com.
• A defined design system and a written product direction.
• Production architecture in build: Cloudflare-first control plane, self-hosted EU voice plane, per-customer isolation, voice layer replaceable.

Next up is pilot readiness for a first grounded voice line at a flagship live event across the 2026 season.

Festivals and complex live events are the beachhead — the sharpest version of the problem, not the whole company. The same grounded voice serves any organisation with a duty of care to a public it cannot currently reach well by phone. Phone is the first channel and the reliability spine; additional surfaces extend the same runtime rather than replace it. Booth stays a managed product, not a platform.

Henki Papp, founder. Senior systems engineer at i3D.net (a Ubisoft company), with large-scale Linux experience at Morgan Stanley before that. 10+ Dutch festivals worked on crew.

Keeping systems up when demand spikes is the day job. Booth is that same problem pointed at a different room.

• Pilot partners. An intro to an organisation running a busy public phone line this season — a festival, venue, or public-facing service.
• Accelerators and mentors. Go-to-market guidance and a founder network.
• Early angels. A relationship now, ahead of a pre-seed round around Q4 2026 or Q1 2027.