Safety and privacy
The AI reads the road. An engineer makes the call.
Traffic signals affect everyone on the road, so the rules come first: who decides, what the system can't do, and what it never collects.
Safety
- Human approval: a city engineer reviews every recommendation, and every approval is logged
- Fail-safe: if a feed fails or confidence is low, the system reverts to the existing fixed-time plan
- Authority: any live signal control requires city authorization. Emergency overrides are tested only in the sandbox during the pilot
Every plan also stays inside hard limits:
- Minimum green time on every approach
- Fixed yellow and all-red intervals
- Enough time for pedestrians to cross
- A maximum wait, so no lane gets starved
Privacy
- No license plate reading and no face identification
- Faces and plates are blurred
- No raw footage kept beyond what the pilot needs
- We follow the Data Privacy Act
- We get permission before using any camera feed
Limits we're open about
| Limit | What we do about it |
|---|---|
| No access to real signal controllers | Advisory and simulated control in the pilot, live control as a later phase |
| City data may not be available | We collect our own baseline counts and footage |
| A language model is unsuitable for real-time control | Gemini perceives and explains, and a deterministic optimizer sets timings |
| Camera permissions | We use mock or recorded footage with permission, and say so |
| Accuracy drops at night or in rain | We report accuracy by condition and show confidence |
| Traffic reduction can’t be proven in the timeframe | We present it as a simulated result and a pilot hypothesis |
Keep reading
See it for yourself.
The app runs on sample data. Open the demo account and try the dashboard, the sandbox and the assistant.