A few cities have "talking statues": a QR code on a plinth plays a recorded monologue, written by hand for one landmark. Audio guides and map apps tell you facts about famous places only. Talking Buildings gives a first-person voice to any building anywhere, generated on the spot from open data, so the corner post office and the old cinema talk too, and they talk to each other.
The data it runs on
OpenStreetMap: the building footprint you are pointing at (matched from GPS position plus compass heading), and its tags: building, name, start_date, architect, building:levels, heritage, amenity, wikidata. Accessed via the Overpass API, cached per city tile.
Wikidata: when OSM carries a wikidata tag, we read the item: P31 instance of, P571 inception, P1619 opening date, P84 architect, P149 architectural style, P1435 heritage designation, P5817 state of use.
Wikipedia: the history section of the linked article, summarised into dated events with citations.
Limits and gaps: most ordinary buildings have only a footprint and a building type. Dates are missing for most. Coverage is strong in Europe and patchy in parts of Africa and Asia.
Fallback: with tags only, the building speaks in "generic voice": it says what buildings of its type and street usually saw, labels every line as general, and invites locals to add a memory or a date, which becomes new data.
1. The mission
Every street becomes a library you can talk to. We want anyone, anywhere, to stop on a pavement and hear the life story of the building in front of them, told honestly, with its sources showing.
2. Why people can't resist it
Personification. A building that says "I" and remembers is instantly lovable. People talk back.
Collector's itch. Every building is a card. Whole streets, stations and cinemas become sets to complete.
Local pride. Your own block finally has a voice, and you can add the memory it is missing.
3. How it works for real
Point: GPS plus compass picks candidate footprints; a vision model (Gemini Flash or GPT-4o mini via OpenRouter) checks the photo against the candidates to pick the right one.
Gather: Overpass for OSM tags, the Wikidata API for properties, the Wikipedia REST API for the history section.
Ground: a fast cheap model (DeepSeek or Gemini Flash) turns the sources into a dated fact list, each line tagged with its source and "verified" or "general".
Voice: a frontier LLM (Claude or GPT) speaks as the building, answering only from the fact list, with personality from type and age.
Neighbours: two fact lists go into one prompt that writes a short dialogue between adjacent buildings.
Speak: TTS (ElevenLabs or OpenAI TTS) with a voice picked by building age and type.
4. Cost per use
About $0.01 per conversation: a grounding pass is cached per building forever, then each chat turn is a short call. Voice adds about $0.01 per minute.
Free: 10 buildings a day, text chat, the album. Walker ($3 a month): unlimited buildings, voices, neighbour dialogues and offline city packs.
5. Growth loop
Every building card is shareable with its best line. Locals who add a missing date or memory get credited inside that building's story, and send it to their family. Schools and walking groups run "collect the street" hunts.
6. Risks and guardrails
Truth: facts come only from cited sources; general lines are labelled general. No invented events or people.
Private homes: residential buildings speak only about era and style, never about who lives there.
Hard history: wartime and colonial chapters are told plainly and respectfully, never played for laughs.
Safety: no use while walking into traffic; voice mode works with the phone in a pocket.
7. MVP build order
Overpass lookup by GPS plus a list of nearby buildings to tap (skip vision at first).
Wikidata and Wikipedia fetch, plus a grounding prompt that outputs a dated fact list as JSON.