Blue Plaques
London's historical figures, brought to life through AI.
The brief I gave myself
A LinkedIn post about Google’s Genie hit 30,000 views with one question: what if AI could bring history to life? I wanted to know what that actually felt like — not as a tech demo, but as a place. London’s 1,000+ Blue Plaques are quiet markers most people walk past. What if they spoke?
Three figures, three modern problems
The MVP features three Blue Plaque holders chosen because their concerns map cleanly onto today’s:
- Alan Turing — AI in warfare and the ethics of computational decision-making
- John Maynard Keynes — the macroeconomics of widespread automation
- Bertrand Russell — existential risk, civilizational stakes
Each delivers a 15–20 second monologue, shot 9:16 portrait, single locked-off frame, BBC interview tone, no music. A text question card precedes the monologue. The plaque itself fades in as a chyron.
The constraint that shaped everything
A hard ceiling of $200–300/month in AI costs. With 2,000 hoped-for installs, that’s roughly 30 cloud requests per user per month. The architecture follows directly:
- On-device Gemma via
llama.rnfor fallback and offline use - A Cloudflare Worker proxy in front of Anthropic, Gemini, and OpenAI for cost-capped routing
- Bring-your-own-key flows for power users
- iCloud Key-Value Store for sync — no accounts, no auth server, no Postgres
The cost ceiling isn’t a footnote, it’s the spec. Every architectural decision flows from it.
Where it is now
v1.5 release-prep — cost-control infrastructure, BYO keys, iCloud sync, PostHog, onboarding, App Store metadata. v2.0 backlog includes depth-parallax 3D portraits, a plaque connections graph, ElevenLabs premium narration, and an Android port.