Tessera
Drop a photo of a place: it finds where the photographer stood, the heading and the lens, then your group plans its own version in a shared 3D set.
Built with a teammate at HackUTD 2026; took 2nd place in the Spacetime track. Drop a photo of a place and Tessera finds where the photographer stood, which way they faced and which lens they used, in phone terms such as "2.2x on your iPhone". The solver matches the photo against a 3D model of the spot built from 155 Creative Commons photos: of 11 held-out photos it placed 10, 9 of them within 25 m of their GPS, with a median focal length error of 3%, and it says "Couldn't place" instead of guessing. The place then becomes a shared 3D set. Every person, prop, light and camera is a row in one SpacetimeDB module, so a drag on one screen shows on every other, and Nemotron places objects from a sentence. The group can plan from a browser, a phone or a Telegram chat. A LoRA fine-tune of Nemotron Nano 12B VL reads the lens of a photo. One spot, DUMBO in Brooklyn, is mapped exactly; anywhere else the place is marked approximate. The hosted demo ran until November 11, 2026; the video shows it and the code runs locally.
- Status: Shipped, built November 7, 2026 to November 8, 2026
- Kind: Vision, AI agents
- Event: HackUTD 2026
- Prizes: Spacetime track, 2nd place
- Stack: Python, TypeScript, SpacetimeDB, three.js, Nemotron
- Devpost: https://devpost.com/software/bean-4970h3
- Code: https://github.com/benz16107/tessera
- Video: https://www.youtube.com/watch?v=KxeUG-_GlKs




