A Spy Satellite Simulator in Your Browser, Fueled Entirely by Public Signals
A fully functional, near-zero-cost geospatial intelligence dashboard now runs in a browser on public signals alone. It demonstrates that the bottleneck in open-source intelligence is not data availability but integration, and that a single developer with disciplined engineering can ship a product that looks like a classified system while respecting every upstream license and privacy boundary.
God's Eye View is an open-source browser application that renders a photoreal 3D Earth overlaid with 13 live data layers, including ADS-B flight positions, AIS vessel tracks, satellite orbits computed client-side via SGP4, USGS earthquakes, NASA fire detections, and municipal CCTV feeds projected into 3D cityscapes. Ten of those layers work without any API key or registration. The project, by Bilawal Sidhu, gained 11.9k GitHub stars in two months without a backend cluster or paid data subscriptions—every signal comes from openly broadcast sources.
The realism comes from five counterintuitive engineering choices: the entire scene renders one polling cycle behind real time so discrete snapshots can be interpolated smoothly; aircraft icons use per-frame screen-space heading projection to stay world-stable; satellite positions are propagated in the browser from TLE orbital elements; a geoid-corrected vertical datum snaps objects to actual terrain height; and a render governor with explicit hold/release quotas stops the render loop when idle. A 266 KB living document, CURRENT-STATE.md, records every design decision and its reversal history, while paired policy modules and self-biting QA scripts enforce correctness.
Cost control is equally systematic: all API keys stay server-side behind a local dev proxy, cache and budget governors cap third-party calls, and the default localhost binding keeps keys off the network. The voice agent exposes 28 tools constrained to confirm only successful operations and must pull live scene context before answering. The project's explicit boundary is that it models events, equipment, and infrastructure—not people—and pull requests that cross that line are rejected.
The deliberate one-cycle rendering lag is the single most important design decision in the project: it trades 15–30 seconds of staleness for continuous, interpolated motion, which is what makes the scene read as real rather than as a dashboard.
Treating error direction as asymmetric—always bias visible-but-slightly-high rather than risk burying an object underground—is a rendering philosophy rarely documented, yet it governs every fallback in the vertical datum subsystem.
The project's license engineering is more instructive than its code: CC BY-NC-SA data is packaged separately with deletion instructions, ODbL obligations are scoped to data only, and privacy transforms strip PII from bundled datasets without altering geometry.
A performance document that openly publishes degradation points (34–35 fps at 8,100 targets, 22 fps in 10,000-vessel AIS scenes) with exact measurement conditions builds more trust than any marketing claim of speed.
The voice layer's design—tools locked by sha256 in unit tests, mandatory scene context fetch before answering, and a hard rule against confirming failed operations—treats the agent as a constrained instrument rather than a conversational chatbot, which is the safer pattern for tool-use agents.