home.social

Search

8 results for “mmornati”

  1. Follow-up to my Jev × Home Assistant post: two comments on the last one deserved a real answer, not just a reply.

    @oneclickclaw_io asked why not feed Jev's context to a small local model instead. Fair point - my prompts really were just snapshots (humidity *now*, flow *now*), no memory of what led there.

    @nocalla asked whether that's not exactly what the Bayesian sensor is for.

    Both got tested, not just answered:

    → Why not local: the box is an N100 mini PC, 4 cores, 16 GB RAM, no GPU, already running HA Core + recorder + Zigbee2MQTT. A local model's own numbers (from the router benchmark) put it at 1.5-3s/call on that class of hardware, several times an hour, contending with real-time Zigbee. And the actual gap was never model intelligence - Jev has a 32k-token window and was only getting ~565 tokens of *current* facts.

    → So instead: a compact-history script built on Home Assistant's own recorder statistics, a 24h house timeline, per-room humidity baselines, a broken-sensor safety net (turns out two humidity sensors had been stuck at 0% - again), per-decision shadow/active switches, and a feedback loop - "who was right?" on the phone, scored on a scoreboard, summarised every week.

    → Bayesian sensor: close, but not the same tool. You hand-estimate every probability yourself, it's strictly on/off, treats correlated clues as independent, and has no idea when its own inputs are lying. Jev takes plain facts, returns a full distribution over N options, and reasons over the whole state. Bayesian still wins for signals you already understand and want fully offline (occupancy, say) - and the two combine nicely.

    Real production numbers from the first days: 267 calls, ~206k tokens, $0.0087 that day. And a live finding I didn't expect: the VMC decision disagrees with its own rule 14/14 times at low confidence, while the shutters decision agrees 78/78 times - exactly why the per-decision switch earns its keep.

    Full write-up:
    blog.mornati.net/jev-home-assi

    #HomeAssistant #SmartHome #HomeAutomation #AI #LLM #SelfHosted

  2. Follow-up to my Jev × Home Assistant post: two comments on the last one deserved a real answer, not just a reply.

    @oneclickclaw_io asked why not feed Jev's context to a small local model instead. Fair point - my prompts really were just snapshots (humidity *now*, flow *now*), no memory of what led there.

    @nocalla asked whether that's not exactly what the Bayesian sensor is for.

    Both got tested, not just answered:

    → Why not local: the box is an N100 mini PC, 4 cores, 16 GB RAM, no GPU, already running HA Core + recorder + Zigbee2MQTT. A local model's own numbers (from the router benchmark) put it at 1.5-3s/call on that class of hardware, several times an hour, contending with real-time Zigbee. And the actual gap was never model intelligence - Jev has a 32k-token window and was only getting ~565 tokens of *current* facts.

    → So instead: a compact-history script built on Home Assistant's own recorder statistics, a 24h house timeline, per-room humidity baselines, a broken-sensor safety net (turns out two humidity sensors had been stuck at 0% - again), per-decision shadow/active switches, and a feedback loop - "who was right?" on the phone, scored on a scoreboard, summarised every week.

    → Bayesian sensor: close, but not the same tool. You hand-estimate every probability yourself, it's strictly on/off, treats correlated clues as independent, and has no idea when its own inputs are lying. Jev takes plain facts, returns a full distribution over N options, and reasons over the whole state. Bayesian still wins for signals you already understand and want fully offline (occupancy, say) - and the two combine nicely.

    Real production numbers from the first days: 267 calls, ~206k tokens, $0.0087 that day. And a live finding I didn't expect: the VMC decision disagrees with its own rule 14/14 times at low confidence, while the shutters decision agrees 78/78 times - exactly why the per-decision switch earns its keep.

    Full write-up:
    blog.mornati.net/jev-home-assi

    #HomeAssistant #SmartHome #HomeAutomation #AI #LLM #SelfHosted

  3. Follow-up to my Jev × Home Assistant post: two comments on the last one deserved a real answer, not just a reply.

    @oneclickclaw_io asked why not feed Jev's context to a small local model instead. Fair point - my prompts really were just snapshots (humidity *now*, flow *now*), no memory of what led there.

    @nocalla asked whether that's not exactly what the Bayesian sensor is for.

    Both got tested, not just answered:

    → Why not local: the box is an N100 mini PC, 4 cores, 16 GB RAM, no GPU, already running HA Core + recorder + Zigbee2MQTT. A local model's own numbers (from the router benchmark) put it at 1.5-3s/call on that class of hardware, several times an hour, contending with real-time Zigbee. And the actual gap was never model intelligence - Jev has a 32k-token window and was only getting ~565 tokens of *current* facts.

    → So instead: a compact-history script built on Home Assistant's own recorder statistics, a 24h house timeline, per-room humidity baselines, a broken-sensor safety net (turns out two humidity sensors had been stuck at 0% - again), per-decision shadow/active switches, and a feedback loop - "who was right?" on the phone, scored on a scoreboard, summarised every week.

    → Bayesian sensor: close, but not the same tool. You hand-estimate every probability yourself, it's strictly on/off, treats correlated clues as independent, and has no idea when its own inputs are lying. Jev takes plain facts, returns a full distribution over N options, and reasons over the whole state. Bayesian still wins for signals you already understand and want fully offline (occupancy, say) - and the two combine nicely.

    Real production numbers from the first days: 267 calls, ~206k tokens, $0.0087 that day. And a live finding I didn't expect: the VMC decision disagrees with its own rule 14/14 times at low confidence, while the shutters decision agrees 78/78 times - exactly why the per-decision switch earns its keep.

    Full write-up:
    blog.mornati.net/jev-home-assi

  4. My home automations know thresholds, not context.

    Sump pump runs 2 min after a rainy night → alert.
    Long shower → "possible leak".
    Washer pauses to soak → "finished", 20 min too early.

    So I gave the "is this normal?" question to Jev, TypeSafe's decision model. You send it a few lines of facts and a typed question, and it answers with a calibrated probability. No free text, no JSON to parse.

    How it's wired into Home Assistant:
    • The old rule always computes its answer. Jev only advises.
    • One switch: shadow (log only), active, or off.
    • An error, a spent budget or low confidence → the rule wins, so the alert still fires.
    • 6 automations: sump pump, water, temperature drops, shutters, ventilation, laundry.

    Cost: 133 calls and 75k tokens on day one, about $3 a year for the whole house.

    Two lessons that aren't about AI:
    1. Writing the facts for the model made me read them. Two humidity sensors had been stuck at 0 %, and my old rules were using them.
    2. My GitOps sync had been failing silently for 5 weeks. Merged ≠ deployed.

    Also in the post: can Laya do the same job locally, to keep occupancy data at home?

    blog.mornati.net/jev-home-assi

    #HomeAssistant #SmartHome #HomeAutomation #LLM #AI #SelfHosted

  5. My home automations know thresholds, not context.

    Sump pump runs 2 min after a rainy night → alert.
    Long shower → "possible leak".
    Washer pauses to soak → "finished", 20 min too early.

    So I gave the "is this normal?" question to Jev, TypeSafe's decision model. You send it a few lines of facts and a typed question, and it answers with a calibrated probability. No free text, no JSON to parse.

    How it's wired into Home Assistant:
    • The old rule always computes its answer. Jev only advises.
    • One switch: shadow (log only), active, or off.
    • An error, a spent budget or low confidence → the rule wins, so the alert still fires.
    • 6 automations: sump pump, water, temperature drops, shutters, ventilation, laundry.

    Cost: 133 calls and 75k tokens on day one, about $3 a year for the whole house.

    Two lessons that aren't about AI:
    1. Writing the facts for the model made me read them. Two humidity sensors had been stuck at 0 %, and my old rules were using them.
    2. My GitOps sync had been failing silently for 5 weeks. Merged ≠ deployed.

    Also in the post: can Laya do the same job locally, to keep occupancy data at home?

    blog.mornati.net/jev-home-assi

    #HomeAssistant #SmartHome #HomeAutomation #LLM #AI #SelfHosted

  6. My home automations know thresholds, not context.

    Sump pump runs 2 min after a rainy night → alert.
    Long shower → "possible leak".
    Washer pauses to soak → "finished", 20 min too early.

    So I gave the "is this normal?" question to Jev, TypeSafe's decision model. You send it a few lines of facts and a typed question, and it answers with a calibrated probability. No free text, no JSON to parse.

    How it's wired into Home Assistant:
    • The old rule always computes its answer. Jev only advises.
    • One switch: shadow (log only), active, or off.
    • An error, a spent budget or low confidence → the rule wins, so the alert still fires.
    • 6 automations: sump pump, water, temperature drops, shutters, ventilation, laundry.

    Cost: 133 calls and 75k tokens on day one, about $3 a year for the whole house.

    Two lessons that aren't about AI:
    1. Writing the facts for the model made me read them. Two humidity sensors had been stuck at 0 %, and my old rules were using them.
    2. My GitOps sync had been failing silently for 5 weeks. Merged ≠ deployed.

    Also in the post: can Laya do the same job locally, to keep occupancy data at home?

    blog.mornati.net/jev-home-assi

    #HomeAssistant #SmartHome #HomeAutomation #LLM #AI #SelfHosted

  7. My home automations know thresholds, not context.

    Sump pump runs 2 min after a rainy night → alert.
    Long shower → "possible leak".
    Washer pauses to soak → "finished", 20 min too early.

    So I gave the "is this normal?" question to Jev, TypeSafe's decision model. You send it a few lines of facts and a typed question, and it answers with a calibrated probability. No free text, no JSON to parse.

    How it's wired into Home Assistant:
    • The old rule always computes its answer. Jev only advises.
    • One switch: shadow (log only), active, or off.
    • An error, a spent budget or low confidence → the rule wins, so the alert still fires.
    • 6 automations: sump pump, water, temperature drops, shutters, ventilation, laundry.

    Cost: 133 calls and 75k tokens on day one, about $3 a year for the whole house.

    Two lessons that aren't about AI:
    1. Writing the facts for the model made me read them. Two humidity sensors had been stuck at 0 %, and my old rules were using them.
    2. My GitOps sync had been failing silently for 5 weeks. Merged ≠ deployed.

    Also in the post: can Laya do the same job locally, to keep occupancy data at home?

    blog.mornati.net/jev-home-assi

    #HomeAssistant #SmartHome #HomeAutomation #LLM #AI #SelfHosted

  8. My home automations know thresholds, not context.

    Sump pump runs 2 min after a rainy night → alert.
    Long shower → "possible leak".
    Washer pauses to soak → "finished", 20 min too early.

    So I gave the "is this normal?" question to Jev, TypeSafe's decision model. You send it a few lines of facts and a typed question, and it answers with a calibrated probability. No free text, no JSON to parse.

    How it's wired into Home Assistant:
    • The old rule always computes its answer. Jev only advises.
    • One switch: shadow (log only), active, or off.
    • An error, a spent budget or low confidence → the rule wins, so the alert still fires.
    • 6 automations: sump pump, water, temperature drops, shutters, ventilation, laundry.

    Cost: 133 calls and 75k tokens on day one, about $3 a year for the whole house.

    Two lessons that aren't about AI:
    1. Writing the facts for the model made me read them. Two humidity sensors had been stuck at 0 %, and my old rules were using them.
    2. My GitOps sync had been failing silently for 5 weeks. Merged ≠ deployed.

    Also in the post: can Laya do the same job locally, to keep occupancy data at home?

    blog.mornati.net/jev-home-assi