Home Environment Monitor & Weather Assistant
A self-hosted home-environment system that turns several asynchronous inputs into timely, explainable household decisions. Deterministic rules decide what the evidence supports; a small local language model may narrate that result, but it never receives the raw numbers or invents the recommendation.
01 · Problem
What needed solving
Indoor humidity, air quality, ventilation and rapidly changing local weather interact. Separate dashboards show measurements, but they do not answer the practical question: what should I do now, and why?
Combine home automation, environmental data and local AI in a system where advice remains testable and safe when any one service fails.
02 · Constraints
The difficult parts
- Reconcile seven concurrent inputs with different update schedules and failure modes.
- Turn temperature and relative humidity into decisions that reflect absolute humidity and real ventilation conditions.
- Use local language generation without letting it alter measurements, thresholds or safety boundaries.
- Keep the service useful during API, model or sensor outages.
03 · Approach
How the work was structured
- Built one asynchronous service around sensor messages, local radio data, official warnings and forecasts.
- Stored time-series data in InfluxDB and used Grafana for long-term inspection.
- Encoded ventilation, humidity and air-quality decisions as deterministic, bilingual rules shared by live and forecast paths.
- Passed only an approved categorical result—not raw measurements—to a local 1.5B language model for concise narration.
- Rejected generated text containing numbers, advice, health claims or unsupported descriptors and fell back to the deterministic message.
Validation and results
The system has 241 automated tests covering rule behaviour, failure paths and message constraints. It runs continuously under a hardened systemd service with monitored inputs and verified backups.
The working system provides dashboards, warnings, forecasts and command responses through Telegram while preserving a strict boundary between calculation and generated language.
05 · Learning
What the project clarified
- AI is safer and more useful here as a constrained presentation layer, not the decision engine.
- A forecast is easier to trust when it uses the same predicate as the live recommendation.
- Environmental advice needs physical reasoning and explicit failure behaviour, not just threshold alerts.