Know when something needs attention—and who has acknowledged it.

The workflow

The property depends on stored water, and nobody is there to look at the tanks. Two level sensors report when storage drops to half and to a quarter. The failure to avoid is not the sensor tripping. It is the alert landing somewhere nobody is looking, and no one knowing whether anyone has acted.

A second need is routine visibility: a regular picture of a water-level gauge so the trend is visible without driving out.

What I built: tank alerts

• Two thresholds with different urgency. The half-full alert is a warning that something upstream may be leaking. The quarter-full alert is critical and the call language escalates.
• Each alert goes out on independent channels at once: a Discord message to me, an email to a family member, and phone calls to two people at the same time.
• The call asks the recipient to press 1 or say yes. That acknowledgement is recorded, announced in Discord, and stops the repeats.
• If nobody acknowledges, the alert fires again every 15 minutes.
• Calls only happen between 8am and 11pm Pacific. Outside that window the email and Discord message still go, so nobody is woken for a warning-level event but the alert is not lost.
• Each channel is sent independently, so a failure in one cannot block the others.

What I built: the camera

• Scheduled snapshots, with a rolling set of recent images plus one per day and one per month kept as an archive.
• A dashboard shows all three views and has a capture-now button, rate limited so it cannot be hammered.
• A health check runs every two hours and posts to Discord if the newest snapshot is more than eight hours old. I added it after a camera outage of about three days went unnoticed.

A design decision worth explaining

The camera sits on a home network that the cloud cannot reach. My first attempts pushed from the cloud into that network through a VM and later a VPN sidecar, and they kept failing in ways that were hard to diagnose.

I reversed the direction. A small program on a PC at the property holds one outbound connection to the database. The cloud function writes a capture command; the database pushes it down that connection; the PC takes the photo and uploads it. No inbound ports, no VPN. If two PCs are listening, a transaction lets the first one claim the command so there is no duplicate capture.

The lesson I took from it: when pushing into a private network is unreliable, have the device on the inside ask for work.

What stays manual

• Someone has to answer the call. The system escalates and repeats; it cannot fix a tank.
• I do not claim every path in the design is running. The sensor vendor's direct push integration needs a larger device count than this setup has, so it is not deployed. Alerts arrive through the vendor's alert emails and scheduled sensor checks instead.