Personal

The infrastructure is interesting. The platform is satisfying to build. But none of it pays its rent unless it actually does something for the household. The personal project line is where the homelab proves it earns its keep.

The discipline is consistent across every automation here: nothing depends on a SaaS, nothing depends on a cloud, everything is owned end-to-end, every signal that drives a decision is captured locally and stored locally. If the broadband drops, the bins still get notified. If a vendor pivots, the heating still works. The bar for adding a new automation is that it has to survive a permanent disconnection from the internet.

The North Lanarkshire Council bin schedule is a colour-coded calendar PDF — Wednesday collection, three-week green / blue / black rotation, brown bin every other week. Easy to read; surprisingly fiddly to automate without producing false-positive notifications six days a week.

The shape that landed:

• schedule.json with metadata (address, confirmed-from / confirmed-until dates, rotation rules) and an array of confirmed dates. When today goes past confirmed_until, a projection layer continues the rotation forward up to 90 days, prefixing output with (PROJECTED) so the source of truth stays obvious.
• check_bins.sh as the dashboard sensor — always shows the next collection with a countdown.
• check_tomorrow.sh as the notification sensor — returns “none” six days a week and only returns bin names when tomorrow is collection day. This split was the key fix: previously the dashboard sensor was used for both display and notification, which fired every day.
• check_schedule_expiry.sh for the bell-curve reminder — fires notifications at ±28, ±14, ±7, ±3, and 0 days relative to confirmed_until so the schedule never silently lapses.
• check_address.sh so the dashboard card shows the property the schedule belongs to (useful when the schedule changes after moving house).

All four sensors are HA command_line sensors with sensible scan_intervals. All deployed via Ansible — the files live on iaccn01, the playbook copies them into the HA container’s /config/bin-schedule/ mount. Pushing a new schedule is cat > schedule.json && ansible-playbook home-services.yml.

There’s also a separate Bin.Watch React portal built for the design exercise — iCal parsing, monthly calendar, alerts cross-referenced against Open-Meteo weather forecasts, configurable bin colour mapping. That’s the “could I rebuild the entire UX of bin tracking from scratch as a proper PWA” version; the HA implementation is the one that actually wakes me up on a Tuesday.

The system has three jobs: detect that I’m in bed and asleep, run a wake sequence aligned to 90-minute cycles with weekday / weekend / OOO / sick logic, and never falsely wake me at 04:00 because someone moved the duvet.

The sensor stack is the bit that took the most iteration. Resistive force sensors were rejected (drift under sustained static load). Under-mattress pads were rejected (8Sleep pod cover muddies the signal). The settled answer is layered:

1. 4× load cells under the bed legs, HX711 ADC, ESP32 running ESPHome. Layer-independent — the 8Sleep, topper, sheets, duvet are all irrelevant. Strain gauges are industrially proven for long-term static load. Gives weight delta plus centre-of-mass for posture inference.
2. Seeed MR24BSD1 mmWave at headboard height for localised head-present and breathing confirmation.
3. Wider mmWave under the bed frame (LD2410B class) for whole-body micro-movement and restlessness signals.

The brain isn’t a stack of HA automations; it’s a dedicated Python container publishing and subscribing over MQTT, with a real state machine (SLEEPING → ALARMING → DISMISSED → CONFIRMED_AWAKE, plus SNOOZED, BRIEF_WAKE, COUCH_SLEEP, PUSH_TO_11, AWAY branches). Every event lands in a SQLite log so the same data feeds health analytics downstream.

Future integrations subscribe to the same MQTT topics: smart blinds, smart TRVs for room-level heating, a future Grafana / React health dashboard reading the event log. The brain doesn’t change when peripherals are added.

The operator principle is non-negotiable: ML can surface “you slept 6h on average this week, your wake threshold is high relative to that” — but it cannot autonomously change the threshold. Suggestions in a sidebar; tuning by hand.

1. April 2026

   Home Assistant up on the IaC pipeline

   home-services VM 104 stood up via Terraform + Ansible. HA container on host networking, Mosquitto MQTT broker. The platform that everything personal hangs off.
2. April 2026

   Bin Collection v1

   Schedule JSON, four command-line sensors, projection layer, bell-curve schedule-expiry notifications, dedicated alert sensor pattern.
3. April 2026

   Bin.Watch React portal

   Parallel build exploring the design exercise of "what does bin tracking look like as a real PWA". Calendar view, weather-cross-referenced alerts, configurable colour mapping. Useful design system reference; HA is still the production path.
4. April 2026

   Smart Sleep System architecture v1

   Sensor stack settled (4× load cells + headboard mmWave + bed-frame mmWave), Python state-machine container chosen over HA automations, MQTT topic structure designed, future integrations (blinds, climate) mapped to the same event bus.
5. Next

   Sleep system hardware procurement

   Load cells + HX711 + ESP32 (~£25 total) + MR24BSD1 + LD2410B. Bench-build first, then bedroom install.
6. Future

   Smart TRVs + boiler relay for per-room climate

   Zigbee TRVs (Sonoff TRVZB class) on each radiator, Shelly relay on the boiler. HA → MQTT events from the sleep system drive bedroom cool-down at pre-sleep and warm-up before wake.
7. Future

   Health analytics dashboard

   Grafana or custom React over the SQLite event log. Sleep duration heatmap, bedtime drift, restlessness over time. The point of capturing every event from day one.