Understeer or oversteer
The model either waits with generic advice or overreacts with broad cooling or an unnecessary dispatch.
Illustrative consumer-HVAC service scenario
Before and after: when an HVAC model gains perception
Switch between the evidence a customer-support model can infer from text, photos and public forums, and the bounded, consented physical context that makes its advice current, explainable and personal.
Decision summary
Two outcomes from the same complaint
Sparse evidence
Context-aware control
Placement, airflow, material and heat context support a bounded, explainable recommendation—not an automatic setting change.
Before uses the complaint, a photo and public anecdotes. Use the two decision options below to see how sparse evidence can fail.
Why fixed sensors misread this office
Air currents, local heat and desk obstruction mean none of the fixed sensors is measuring neutral room air.
Read the blue airflow as the HVAC supply and cooler draft; red currents are heat plumes from the fireplace and oven.
Text-only support sees a complaint; the placed sensors disagree, lag and pick up local drafts.
Sensor ASupply draft · quick swings
Sensor BBehind desks · delayed average
Sensor COven plume · hot bias
Cool airflowHVAC and drinks-cooler drafts
Warm airflowFireplace and oven heat plumes
What reaches the HVAC company
The first three inputs are useful, but they cannot establish the unit’s current physical situation.
01
Support chat“Desks feel stuffy after lunch. The system runs for hours.”02
Customer photoOne office snapshot; no reliable sensor position, airflow path or heat-source distance.03
Public forumThree Reddit-like posts mention a similar complaint.04
Wi-Fi fingerprintRelative sensor positions and room zones, with consent.05
Computer visionMaterial class and thermal-retention cues; no identity capture.06
IR proximityDistance and intensity of nearby heat sources.Model reasoning and response
Choose a likely failure mode. The legend below explains both outcomes.
Understeer: send generic guidance and wait, conserving energy but leaving the customer unsupported. Oversteer: treat a local load as a unit fault, pushing broad cooling or dispatch without enough evidence.
Illustrative outcome of the text-only decision
Model output
All values are synthetic examples for discussion. Occupancy, “away” or other customer context must be explicitly provided; the system does not identify people, infer travel or change HVAC settings automatically.