Walk it if you like. Drop a camera roll to bring the evening back.
This page is a working model of the experience. The pieces underneath it are at different levels of maturity, and the pitch should say so.
One controlled walkthrough by staff on a quiet afternoon produces a full 3D capture of the lounge — a Gaussian splat you can walk through in a browser. Luma, Polycam, or open-source nerfstudio do this now. That capture is the stage; it is made once and reused for every evening.
Each member photo is matched against the room capture to recover where the camera was standing and which way it faced. This is visual localization — the same maths behind AR anchoring. It works well in decent light and struggles in the dark, which is the honest constraint for a cigar lounge; the fix is a strong master capture and phones on night mode.
Photos straight from a camera roll carry timestamp, device, and often location. Anything that passed through Messages, WhatsApp, or Instagram arrives stripped. For those, the model looks at the room, the lighting, and the calendar, proposes a night, and asks when it is not sure. It never silently guesses.
What this page shows is photos anchored where they were taken, fading through the night — which already tells the booth's story. Rebuilding the people themselves as 3D figures that move through time is 4D reconstruction, and it is research-grade today. It is the roadmap, not the launch. Say that in the room and the rest of the pitch is believed.
A private club runs on discretion. Every evening is visible only to the members who were there. Face tagging is opt-in per member, anyone can remove their likeness from any photo, and a member can hide a photo they uploaded at any time. Nothing leaves the room.
Uploads go to R2 through the Worker. Metadata extraction, localization, and tagging run as a queue. Claude reads the stripped photos and drafts the question. The viewer is the same Cloudflare front end as everything else, with the room capture served as a static asset.