What failure modes does it catch and how early?

The agent monitors compressor head pressure, fan motor amperage draw, evaporator coil delta-T, condenser approach temperature, and refrigerant circuit indicators. For the most common failure modes — compressor overload, fan motor bearing degradation, evaporator coil fouling — the signal emerges 7–14 days before failure in the majority of cases in our production deployments. Bearing failures on fan motors often show 3–4 weeks out. Sudden failures (refrigerant leak, electrical fault) are caught within hours of the first anomalous reading, but cannot be predicted days in advance by any monitoring system.

How do you suppress false positives?

The primary mechanism is a persistence filter: a parameter must deviate from the asset's baseline consistently across multiple operating cycles and changing outdoor conditions before an alert fires. A compressor that runs harder during a 95°F day is expected — the agent knows the asset's load-temperature curve. If head pressure is elevated on moderate days, over three consecutive operating periods, the signal survives the filter. Secondary suppression comes from correlation: the agent looks for corroborating signals before escalating a single-point anomaly.

What's the typical false-positive rate?

In production deployments, we target and typically achieve a false-positive rate below 8% — meaning fewer than 1 in 12 alerts results in a work order where no actionable fault is found. The rate depends heavily on equipment age and sensor data quality; older units with noisy sensors run slightly higher. We track false-positive rate per asset and per customer, and we tune the persistence filter thresholds during the first 30 days of deployment based on what we observe. If your team's trust in alerts erodes due to false positives, that's a deployment problem we treat as P1.

Does it integrate with our CMMS — UpKeep, Limble, eMaint, or IBM Maximo?

Yes. Production integrations exist for UpKeep, Limble, eMaint, and IBM Maximo. When an alert clears the persistence filter, the agent creates a work order in your CMMS pre-populated with the asset ID, fault category, the specific parameter that triggered it, the RUL estimate, and the recommended action. No manual triage, no copy-paste. If your CMMS isn't on that list, contact us — we integrate with any platform that exposes a REST API or webhook endpoint for work order creation.

How is RUL estimated and how accurate is it?

Remaining Useful Life is estimated from the drift trajectory of the flagged parameter combined with historical failure-time data for the same failure mode and equipment class. For example, a compressor whose head pressure is drifting up at 0.4 PSI per operating day, on a unit with the observed baseline, is compared against similar units that failed — giving a distribution of days-to-failure. The median of that distribution is reported as RUL, with a confidence range. Accuracy is best (±20%) for slow-moving degradation modes like coil fouling. For faster-moving failures, treat RUL as a priority signal rather than a precise date.

How is baseline learning done — per asset or per brand?

Per asset. The agent observes each individual unit for 14–21 days before beginning to flag anomalies — learning its specific load curves, runtime patterns, and normal parameter ranges under different outdoor conditions and occupancy loads. A Carrier RTU on the south face of a building in Phoenix runs differently from the same model on the north face in Chicago, and the agent baselines them separately. Brand-level failure-mode data informs the RUL model, but alert thresholds are asset-specific, not brand averages.

How is this different from our BMS vendor's 'predictive analytics' module?

BMS vendor predictive modules typically apply static threshold alerts — 'head pressure above X PSI' — and report on a single brand's data. They don't learn per-asset baselines, they don't suppres false positives via persistence filtering, and they don't integrate with CMMS platforms outside the vendor's own ecosystem. The HVAC AI Fault Detection agent works cross-brand, learns individual asset baselines, applies persistence filtering to reduce noise, and delivers work orders directly to whichever CMMS your team already uses — without replacing your BMS.

How is fault detection billed?

Per monitored asset (RTU, AHU, chiller, or packaged unit), with tiering by asset count. There's no per-point or per-alert metering — all failure modes, all CMMS integrations, and all RUL reports are included. A 30-day pilot on a representative subset of your fleet is structured as a fixed-fee engagement so you can validate alert quality before committing. Portfolio rollout moves to an annual contract. Contact us with your asset count and CMMS platform for a specific quote.

Private beta · Fault detection

Catch HVAC faults before your tenants — or your auditors — do.

AI agents that learn each building's baseline, watch every unit 24/7, and call the failure 7+ days before it stops you. Alarm noise down; real faults up.

Join the waitlist →

7+ days

early warning on key failures

compressor / fan / economizer

24/7

anomaly watch

every unit, every metric

less noise

more signal

baselines tuned per-building

Three modes — autonomous leads here

Primary for this page

24/7 anomaly watch

Learns each building's normal in the first 2-4 weeks. Then flags only the deviations that matter.

Diagnostic copilot

Ops asks: 'why is RTU-3 flagged?' Copilot summarizes the deviation + suggests the probe.

MCP for your CMMS

Auto-opens tickets in ServiceChannel / Maximo / Fiix via MCP.

Predictive fault detection

7+ days of early warning on compressor, fan, and coil failures — per asset, not per brand average.

HVAC AI Fault Detection learns each asset's individual baseline, flags persistent drift before it becomes a failure, and delivers a Remaining Useful Life estimate to your CMMS — so you schedule the repair, not the emergency truck roll.

Reactive maintenance is expensive in a predictable way: the equipment runs until it fails, the failure happens at the worst possible time, emergency parts cost 3–5× list price, and overtime and rental equipment add up fast. Preventive maintenance trades that cost for a different inefficiency: parts are replaced on a calendar, not on condition, which means most replacements happen before the equipment actually needs them.

Fault Detection sits between those two modes. The agent reads your HVAC telemetry — compressor head pressure, fan motor amperage, evaporator coil delta-T, condenser approach temperature — and learns what normal looks like for this specific unit over the first 14–21 days. That per-asset baseline is more accurate than a brand-average threshold because it accounts for the quirks of your building, your climate, and your equipment's age and configuration.

When a parameter drifts outside the asset's normal envelope, the agent doesn't fire immediately — it waits to see if the drift persists across changing operating conditions. A compressor that runs a little harder during a heat wave is normal. A compressor whose head pressure is 8% above baseline at moderate outdoor temperature on three consecutive operating days is a signal. That persistence filter is the primary mechanism for keeping false-positive rates low.

When a signal survives the filter, the agent computes a Remaining Useful Life estimate based on the drift trajectory and historical failure patterns for the component and brand. The alert routes to your CMMS — UpKeep, Limble, eMaint, or IBM Maximo — as a pre-configured work order with asset context, the trending parameter, the RUL estimate, and the recommended action. Your maintenance team schedules the repair on their terms, not the equipment's.

Where it pays off

Concrete scenarios from asset-intensive operations.

Four patterns we see repeatedly across facilities, manufacturing, and multi-site HVAC portfolios.

Multi-site facilities manager

30-site portfolio, compressor failures hitting 3–4 sites per year without warning. Emergency rental units, overtime, tenant escalations — each event costs $15k–$40k all-in.

→ Fault Detection flags compressor head-pressure drift 7–10 days before failure at two sites in the first quarter. Both are repaired on schedule. Zero emergency truck rolls.

7+ days early warning, 100% of flagged failures caught before breakdown

Predictive maintenance lead

Parts budget is set annually on a calendar-replacement schedule. Over-replacing wastes budget; under-replacing causes failures. RUL data doesn't exist, so conservative schedules win.

→ Per-asset RUL forecasts let the PM lead see which units genuinely need replacement this quarter and which have 18 more months. Parts budget shifts from calendar-driven to condition-driven.

15–25% parts budget reduction vs. conservative calendar schedule

CMMS-integrated maintenance team

Monitoring alerts arrive as emails to a shared inbox. Creating work orders is manual, gets missed on busy days, and loses the telemetry context that triggered the alert.

→ Fault Detection routes every alert directly to UpKeep as a pre-populated work order: asset ID, fault type, trending parameter, RUL estimate, and recommended action. No inbox triage, no manual entry.

Zero missed alerts, work order created in under 60 seconds of detection

Asset-heavy operations (DC or light manufacturing)

Capex planning for HVAC replacement is guesswork — no condition data, so finance gets a replacement list driven by age and gut feel, not actual equipment state.

→ RUL forecasts across the full HVAC fleet let ops deliver a 3-year replacement schedule grounded in condition data. Finance gets a defensible model; ops avoids surprise capital requests.

3-year capex forecast accuracy from ±40% (age-based) to ±15% (RUL-based)

FAQ

Fault detection — common questions.

What failure modes does it catch and how early?

The agent monitors compressor head pressure, fan motor amperage draw, evaporator coil delta-T, condenser approach temperature, and refrigerant circuit indicators. For the most common failure modes — compressor overload, fan motor bearing degradation, evaporator coil fouling — the signal emerges 7–14 days before failure in the majority of cases in our production deployments. Bearing failures on fan motors often show 3–4 weeks out. Sudden failures (refrigerant leak, electrical fault) are caught within hours of the first anomalous reading, but cannot be predicted days in advance by any monitoring system.
How do you suppress false positives?

The primary mechanism is a persistence filter: a parameter must deviate from the asset's baseline consistently across multiple operating cycles and changing outdoor conditions before an alert fires. A compressor that runs harder during a 95°F day is expected — the agent knows the asset's load-temperature curve. If head pressure is elevated on moderate days, over three consecutive operating periods, the signal survives the filter. Secondary suppression comes from correlation: the agent looks for corroborating signals before escalating a single-point anomaly.
What's the typical false-positive rate?

In production deployments, we target and typically achieve a false-positive rate below 8% — meaning fewer than 1 in 12 alerts results in a work order where no actionable fault is found. The rate depends heavily on equipment age and sensor data quality; older units with noisy sensors run slightly higher. We track false-positive rate per asset and per customer, and we tune the persistence filter thresholds during the first 30 days of deployment based on what we observe. If your team's trust in alerts erodes due to false positives, that's a deployment problem we treat as P1.
Does it integrate with our CMMS — UpKeep, Limble, eMaint, or IBM Maximo?

Yes. Production integrations exist for UpKeep, Limble, eMaint, and IBM Maximo. When an alert clears the persistence filter, the agent creates a work order in your CMMS pre-populated with the asset ID, fault category, the specific parameter that triggered it, the RUL estimate, and the recommended action. No manual triage, no copy-paste. If your CMMS isn't on that list, contact us — we integrate with any platform that exposes a REST API or webhook endpoint for work order creation.
How is RUL estimated and how accurate is it?

Remaining Useful Life is estimated from the drift trajectory of the flagged parameter combined with historical failure-time data for the same failure mode and equipment class. For example, a compressor whose head pressure is drifting up at 0.4 PSI per operating day, on a unit with the observed baseline, is compared against similar units that failed — giving a distribution of days-to-failure. The median of that distribution is reported as RUL, with a confidence range. Accuracy is best (±20%) for slow-moving degradation modes like coil fouling. For faster-moving failures, treat RUL as a priority signal rather than a precise date.
How is baseline learning done — per asset or per brand?

Per asset. The agent observes each individual unit for 14–21 days before beginning to flag anomalies — learning its specific load curves, runtime patterns, and normal parameter ranges under different outdoor conditions and occupancy loads. A Carrier RTU on the south face of a building in Phoenix runs differently from the same model on the north face in Chicago, and the agent baselines them separately. Brand-level failure-mode data informs the RUL model, but alert thresholds are asset-specific, not brand averages.
How is this different from our BMS vendor's 'predictive analytics' module?

BMS vendor predictive modules typically apply static threshold alerts — 'head pressure above X PSI' — and report on a single brand's data. They don't learn per-asset baselines, they don't suppres false positives via persistence filtering, and they don't integrate with CMMS platforms outside the vendor's own ecosystem. The HVAC AI Fault Detection agent works cross-brand, learns individual asset baselines, applies persistence filtering to reduce noise, and delivers work orders directly to whichever CMMS your team already uses — without replacing your BMS.
How is fault detection billed?

Per monitored asset (RTU, AHU, chiller, or packaged unit), with tiering by asset count. There's no per-point or per-alert metering — all failure modes, all CMMS integrations, and all RUL reports are included. A 30-day pilot on a representative subset of your fleet is structured as a fixed-fee engagement so you can validate alert quality before committing. Portfolio rollout moves to an annual contract. Contact us with your asset count and CMMS platform for a specific quote.

Still have a question?

Ask the fault-detection agent.

Type a question. We'll route it to the right page or capture it for our team.

What failure modes does it catch? How accurate is RUL? Integrate with UpKeep or Maximo? Alarm diagnosis for techs? Cut energy spend 20–30%?

Speaks to your existing kit

Carrier, Trane, Daikin, Mitsubishi, LG, Lennox, York, Samsung — 20+ HVAC, home-automation, and BMS brands.

63 brands across 3 categories — HVAC (31), Home Automation (18), BMS (14). Protocols: BACnet, KNX, MQTT, Matter, Modbus, REST, WebSocket, Z-Wave, Zigbee.

63

See full coverage →

How it stays out of your way

Secure

Sealed data plane. Per-site auth. Audit log on every setpoint touch.

Runs on the edge

Deploys at the building edge — your data doesn't leave the site to be useful.

BYO LLM

Works with Claude, ChatGPT, and any MCP-compatible client. You pick the brain.

Private beta

Get the next 30 days of failures ahead of time.

Designed for building operations + predictive-maintenance teams.