Architecture
How FrostLogic Explore works
Edge Agent on site, detection and ranking in the cloud, write-back only inside scopes you grant. This page is the method end to end.
- Writes only where you allow
- Seven detector types
- EU-hosted app
The path
From the BMS to a ranked decision
Eight stages. Same product on every building that connects.
01
Edge Agent
On-prem beside the BMS. Outbound HTTPS only.
02
Data
Points land in Explore. Sources reconciled. Defaults per source.
03
Seven detectors
Statistical outlier through setpoint tracking.
04
AI investigations
Related anomalies grouped. Verdicts with evidence.
05
Attention Queue
One ranked list of what needs a human.
06
Forecasts
Nested 50% and 90% bands, refreshed intraday.
07
Permissioned write-back
Read-only until you grant a scope.
08
Smart Schedules
Available in pilot. Default off. LLM never in the control loop.
1 · Ingest
FrostLogic Edge Agent beside the BMS
The FrostLogic Edge Agent reads your BMS, meters and sensors, and writes back to the points you have granted it. It runs on a Windows machine on site. Traffic is outbound HTTPS only; nothing inbound is required to keep the feed alive.
Protocols supported today: BACnet, OPC UA, Modbus, MQTT, M-Bus, oBIX. Vendor estates we already connect include Siemens Desigo CC, Kieback&Peter, Regin, Bastec, and the Niagara / Tridium family.
Buildings on Schneider EcoStruxure Building Data Platform (BDP) can take a cloud path instead of, or alongside, the local agent. Either way, points arrive in the same Explore model.
2 · Model
Points land once, then stay coherent
Ingested points are reconciled against source identity so the same meter does not appear twice under two names. Detection and forecast defaults attach per source, then operators can tighten them.
The durable store sits under the data warehouse capability. This page is the path into that store, not a second warehouse product.
3 · Detect
Seven anomaly detector types
Explore runs seven detector types on the points you connect. Cadence is about every two hours. Pipeline filters cut noise before anything reaches a person: causal filtering via FrostDynamics, a forecast quality gate, and a history-coverage check so short or status-only series do not flood the queue.
Statistical outlier
Values that sit outside the building's own recent distribution.
Energy spike
Sudden load jumps that do not match occupancy or weather.
Forecast deviation
Live series drifting off the multivariate forecast envelope.
Sensor calibration
Drift, stuck, or offset sensors that would otherwise poison control and reporting.
Equipment health
Multi-metric patterns plus degradation that show plant heading the wrong way.
Data gap
Missing intervals that break continuity for detection and compliance evidence.
Setpoint tracking
Plant that fails to hold the setpoint it was told to hold.
4 · Investigate
AI investigations group related anomalies
Related anomalies are clustered into one investigation. Each investigation carries a verdict and the evidence trail back to the signals.
Benign clusters can close on their own. Real issues become actionable insights in the Attention Queue. When degradation is clear, the investigation can flag that failure is looming; it does not invent a fixed N-days-to-failure claim.
5 · Rank
Attention Queue: one home screen
The home screen is one ranked list of what needs a human: insights, anomalies, compliance items, automation approvals, data health, and savings opportunities.
Five personas (facility, energy, sustainability, portfolio, data engineer) change order, not permissions. Same findings, different first screen.
The dashboard is the question. The queue is the answer.
6 · Forecast
Nested forecast bands, refreshed during the day
Forecasts ship with nested 50% and 90% bands. They are multivariate, with horizons at 24 hours, 7 days and 30 days, and they refresh during the day as new points land.
Continuous meter and sensor monitoring that feeds these forecasts lives on energy monitoring.
7 · Act
Permissioned write-back
Explore starts read-only. You grant write access one scope at a time, and you decide whether each change waits for a person or runs on its own.
Spec: write access is off by default, granted per scope. Every new writable point starts disallowed. Unapproved proposals expire. Transitions are audited.
Humans can only tighten bounds. Temporary overrides hand control back when they end. An In force now view shows held points. Writes go through the Edge Agent and, where opted in, Schneider BDP.
This is Sensor Intelligence on top of your BMS, not a BAS replacement and not a CMMS. Integrator logic stays. Explore writes points only inside the leash you set. Detail for the action layer: Automation.
8 · Schedules
Smart Schedules (pilot, default off)
Smart Schedules are available in a default-off pilot. Template library, human approval per bundle, shadow then active.
Architectural proof point: the LLM is never in the control loop. Approved schedules keep running if every AI provider is down. Do not treat this as a standard on-by-default feature.
Topology
What runs where
Edge on site. Application in the EU cloud. FrostDynamics stays supervisory: proposals and approved schedules, not continuous MPC in the control loop.
Trust
Data residency and processing
The application is hosted in the EU on Hetzner. GDPR framing applies to how we process personal data in the product path.
EU-only LLM processing is available on request. That is not the default for every model call. For the fuller trust posture, see Security.
Before you ask
Architecture questions
On a Windows machine on site, beside the BMS. It talks outbound HTTPS only. Writes, when you grant them, go back through that agent to the points inside the scope.
Explore starts read-only. You grant write access one scope at a time, and you decide whether each change waits for a person or runs on its own. Points outside a granted scope stay untouched.
BACnet, OPC UA, Modbus, MQTT, M-Bus and oBIX. Vendor estates already in the field include Siemens Desigo CC, Kieback&Peter, Regin, Bastec and Niagara / Tridium. Schneider BDP is the cloud path where the building already uses it.
The home screen: one ranked list of what needs a human (insights, anomalies, compliance, automation approvals, data health, savings). Personas reorder the list; they do not change who is allowed to act.
The Explore application runs in the EU on Hetzner. EU-only LLM processing is available on request; it is not claimed as the default for every model call. Details live on Security.
Smart Schedules are a default-off pilot. A human approves each bundle. Inside a granted write scope, approved schedules can execute without putting the LLM in the control loop; if AI providers go down, those schedules keep running.
Check one building. Keep the findings.
Connect one building for a free look at what Explore returns. If you want the longer path after that, the 90-day pilot is available. No start fee, no hardware. Prefer plans first? Pricing is public.