ARMS Recommendation Engine

What it does

The engine ingests building structures in the form of a standardized ontology, occupancy data and booking data to generate recommendations that are ranked using scores and confidence metrics. Examples include:

• "This room is underused, reallocate or consolidate."
• "This zone is over capacity at certain hours, reduce HVAC/maintenance costs by rescheduling."
• "This room type is systematically underbooked, investigate why."

Each recommendation carries a subject_key, a target-scoped identity that attaches feedback and later runs back to the same underlying finding across re-runs, rather than treating every run as a clean slate.

Personas were integrated into the engine such that the program creates recommendations based on the type of site/building and the type of person inspecting the recommendations. A healthcare persona would be recommended ways to optimize clinic usage by observing booked clinic space that goes unused, whereas office building managers would receive recommendations on underused office spaces and underperforming meeting rooms based on available amenities.

Design philosophy

To optimize the project for extensibility and a business-logic-first approach, the application was prototyped using three different models: LLMs, traditional TensorFlow-based ML, and a basic statistical model. A statistical model was arrived at, with the ability to extend functionality using TensorFlow and LLMs built in.

Each recommendation type links to a candidate builder, which in its current form runs a statistical model to generate relevant features and metrics for the recommendation. TensorFlow is used for trend prediction and pattern recognition as needed, while still falling back on the statistical model to parse business logic and policies.