Touch the exact place.
The selected latitude and longitude—not a capital city or country centroid—drive the weather request.
Geographic truthThe planet, translated
Touch any point on Earth. Weather Atlas turns precise local conditions into clear, locally relevant language—then speaks the forecast aloud.
Drag to rotate · Tap any point
6.2442° N · 75.5812° W
19°
Light rain
Feels like 19°
En Medellín, la temperatura es de diecinueve grados con lluvia ligera.
A new way to read the world
The interface feels simple because the underlying decisions are explicit. Geography, weather, language, and speech remain aligned from the first click to the final syllable.
The selected latitude and longitude—not a capital city or country centroid—drive the weather request.
Geographic truthCLDR-derived suggestions surface locally relevant languages while preserving the user’s explicit control.
Language without assumptionsThe displayed transcript becomes the exact speech input, reducing drift between source data, text, and narration.
One source of truthThe same exact coordinate anchors the visual selection and the weather lookup.
Provider output is normalized into a bounded schema before any language generation.
The visible transcript remains inspectable, copyable, and accessible before playback.
Precision by design
Weather Atlas is designed around semantic consistency. The globe is not decoration. The transcript is not an afterthought. The voice does not invent the forecast.
Large countries, islands, coastlines, and open ocean remain geographically meaningful.
A region may suggest languages, but the individual decides what the application speaks.
Every spoken result has a visible, reviewable text equivalent and an explicit playback action.
Built as a controlled pipeline
The system reduces provider output to structured facts, localizes only those facts, and synthesizes exactly the transcript the user can see.
Exact coordinates and geographic context.
Bounded canonical facts with source and time.
Constrained transcript in the selected locale.
Server-side speech generation from visible text.
Secrets, caching, functions, and fast delivery.
Search, text, and device speech remain available.
The critical design choice: the model never receives an unbounded weather-provider payload. It receives a small canonical fact set. The transcript it returns is then used as the exact TTS input.
More than a globe demo
The experience includes the guardrails, fallbacks, and operational seams needed to move beyond a visual prototype.
OpenAI credentials stay inside Cloudflare Pages Functions, never in browser code.
Abort controllers and selection IDs prevent stale weather or audio from replacing a newer choice.
Search, visible transcripts, keyboard controls, live regions, and reduced motion work without globe interaction.
Weather, mapping, localization, and speech boundaries can evolve without rewriting the product surface.
Ready for first light
The complete project is packaged for Cloudflare Pages with a Vite + TypeScript client, Pages Functions, generated geographic data, deployment guidance, and production hardening notes.