Be Unorthodox

Case study

AI Pocket Tour Guide

A multimodal pocket tour guide, built solo in a hotel room on a work trip. Point your phone at a building and hear a cloned local-leader voice tell you what you're looking at.

Prototype. A 2024 proof of concept. It ran on my own phone and was never launched.

My role
Sole builder. Concept, the three-API harness, prompts, and voice cloning.
Timeframe
Jun 2024
Built with
AImultimodalElevenLabs

June 2024. The team flew to Toronto for a company on-site at Thomson Reuters HQ. One night, back at the hotel, I built a tour guide instead of going to bed.

The idea was simple. You’re in a new city with your team. You point your phone at a building, a statue, a plaque. A voice tells you what it is. Not a generic narrator. The voice of someone on your own team who actually knows the city.

The harness

A harness, the way I use the word, is several models wired into one experience. Not a chatbot. This one wrapped three APIs:

  • A vision model read the photo and worked out what was in frame.
  • ChatGPT took that subject, researched it, and wrote a short tour-guide script.
  • ElevenLabs spoke the script in a cloned voice of a local leader.

The interesting part is never a single call. It’s the seams. A photo has to become a subject. A subject has to become a script that sounds spoken, not written. The script has to come out in one specific person’s voice and still feel like a tour, not a robot reading an encyclopedia entry aloud. Every handoff is where it breaks, and the handoffs are where the actual work lives.

Why a cloned voice

I could have used a stock narrator and saved myself the hardest third of the build. I didn’t, because the voice is the whole point.

A stranger reading facts about a building is a podcast. Your own colleague’s voice explaining the thing outside your hotel window is a different feeling entirely. It turns a lookup into something that belongs to the trip you’re on. The gimmick is what makes it land, so the gimmick got the effort.

What it actually was

I’ll be straight about the scope. This was a hotel-room build, a proof of concept, n=1. It ran end to end on my own phone that week and it was genuinely fun to use. I never hardened it, never put it in an app store, never rolled it out to the team. It exists to prove the seam between vision, language, and a cloned voice can be stitched into one tap-and-listen experience by one person in a few spare hours.

That’s the reason it’s here. Most of what I build that matters is live and load-bearing. This one is the opposite, and useful for a different reason: it’s the clearest small example of how I think about AI. Find three capabilities that don’t normally talk to each other, and make the seams between them disappear.