Atlas

Inspiration

Matt and I kept coming back to one idea: our current prosperity did not appear overnight. It is the sum of countless moments, places, and people across history.
Most people can read about history, but very few can feel what a place and time might have been like. We wanted to build a way to step into those moments and make history feel alive, personal, and accessible.

What we built

We built an AI-guided exploration experience:

Pick any country, province, city or location on a 3D globe
Talk to an AI guide (through voice) to choose a year and area of the world you want to explore
Warp to hyperspace while the world is generating, with listening to music tailored to the selected time, place, and culture
Arrive inside a rendered 3D world generated for your exact coordinates and chosen moment in history
Explore the world while your guide interprets what you’re seeing and shares the historical context behind it.

The full experience stays inside our app. No redirect is required to explore.

How we built it

Frontend: Vite + React + TypeScript + Three.js + Spark.js
Globe phase: interactive Earth selection for location + year (with infinite zoom-in)
Voice loop (Gradium STT/TTS): low-latency, interruptible conversation so users can cut in at any time and still have a natural back-and-forth
Core logic (Gemini API): Used as the decision layer for guide behavior, prompts, and responses across every phase
Gemini tool calling: drives globe suggestions, era-aware music, and world creation triggers so the map, audio, and generation pipeline stay in sync
World generation: World Labs API with prompts that anchor the location, year, cultural context and users interests deteremined from the conversation.
World rendering: Spark.js + Three.js Gaussian splat rendering pipeline
UX pipeline: starfield + globe -> hyperspace -> loading narration -> exploration

Challenges we faced

Managing conversation state during interrupts: if VAD split one user thought into multiple turns, Gemini could fire twice and create overlapping responses
Making WebSocket streaming robust on the frontend: without strict queueing and response-state control, audio/text events could previously arrive out of order and feel jumpy or buggy
World renderer: splat artifact rendering issues in Spark.js integration (still working on fixing this)
Matching color and visual quality to the reference World Labs viewer

What we learned

Adding strong state transitions (landing page -> globe -> worldview) made the journey more immersive
We found there was a tradeoff between very low STT latency and system stability
Learned to use tool calling much more effectively for structured actions like location selection and dynamic messaging.
Orchestrating voice, world generation, and rendering required robust state management with graceful recovery paths.
The best loading experience is still a meaningful experience: context-aware music and messages keep users inspired.
Experiencing it ourselves, showed us just how special this is and how it creates a new way to learn history and culture.