Light don go
A local AI health reminder tool designed for power-outage scenarios. When a laptop switches to battery power, it triggers a full-screen prompt recommending outdoor activities. The tool operates entirely offline, using a local Gemma model to generate suggestion text.
Timeline (last 90 days)
Light don go is an open-source tool that displays a full-screen page when a laptop switches to battery power for more than 30 seconds, offering suggestions, a 60-second stretching timer, and a close button.
Light don go runs entirely without network access, because when the power goes out the router also loses power, making cloud APIs unreachable.
The tool saves Open-Meteo hourly weather forecasts every 30 minutes when power is on, and reads only the local cache file after a power outage.
The author divides AI responsibilities as follows: code is responsible for generating factual lines containing numbers, while Gemma is only responsible for generating advisory sentences containing no numbers whatsoever.
After splitting the task and using a guard to validate advisory sentences (rejecting outputs containing numbers, exceeding 20 words, or mentioning NEPA/lights/electricity/generators), all 9 tests passed on the first attempt.
The guard only checks for numbers and topics, not semantic coherence, and Gemma once produced a strange sentence like 'watch the rain inside your phone' that passed validation.
Choosing a small model means occasionally odd outputs, but larger models cannot run on the laptop, and the author accepted this tradeoff.
The presentation layer is an HTML file opened by Edge with the --app --start-fullscreen flags; model-generated sentences are injected as data rather than inserted as HTML.
The entire project has 98 tests, one of which blocks all network sockets to verify the power-outage flow in a fully offline state.
All Facts (9)
Light don go is an open-source tool that displays a full-screen page when a laptop switches to battery power for more than 30 seconds, offering suggestions, a 60-second stretching timer, and a close button.
50%UnverifiedLight don go runs entirely without network access, because when the power goes out the router also loses power, making cloud APIs unreachable.
50%UnverifiedThe tool saves Open-Meteo hourly weather forecasts every 30 minutes when power is on, and reads only the local cache file after a power outage.
50%UnverifiedThe author divides AI responsibilities as follows: code is responsible for generating factual lines containing numbers, while Gemma is only responsible for generating advisory sentences containing no numbers whatsoever.
50%UnverifiedAfter splitting the task and using a guard to validate advisory sentences (rejecting outputs containing numbers, exceeding 20 words, or mentioning NEPA/lights/electricity/generators), all 9 tests passed on the first attempt.
50%UnverifiedThe guard only checks for numbers and topics, not semantic coherence, and Gemma once produced a strange sentence like 'watch the rain inside your phone' that passed validation.
50%UnverifiedChoosing a small model means occasionally odd outputs, but larger models cannot run on the laptop, and the author accepted this tradeoff.
50%UnverifiedThe presentation layer is an HTML file opened by Edge with the --app --start-fullscreen flags; model-generated sentences are injected as data rather than inserted as HTML.
50%UnverifiedThe entire project has 98 tests, one of which blocks all network sockets to verify the power-outage flow in a fully offline state.
50%