FW Thorpe (Thorlux)

FW Thorpe runs SmartScan, one of the largest connected lighting estates in the world: roughly one million luminaires across 661 buildings, every fitting carrying a passive infrared occupancy sensor and feeding into Tiger DB on PostgreSQL with TimescaleDB. The data is rich, the platform is mature. What was missing was a business-facing artefact on top of it.

Universities and large corporates make capital decisions on manual space utilisation analysis that takes two to four hours per site. Thorlux cannot scale that across a multi-thousand-site customer base. A single engineering dashboard cannot serve a board member, a facilities manager, and a department head equally well, and customer occupancy data has to stay inside UK jurisdiction, which rules out any default-to-OpenAI architecture.

• Customer occupancy data stays inside UK jurisdiction, end to end.
• Sensor data is the means, not the artefact; the buyer wants the recommendation.
• One source of truth, three audiences, three differently-shaped reports.
• Model selection backed by benchmarking, not vendor preference.
• Architecture sits inside the existing SmartScan deployment, no new data plane.

We recommended dropping the single-dashboard ambition and producing three reports instead, shaped as a pyramid. A one-page executive summary for the C-suite, scoped to underutilised spaces because that is the call the reader is paid to make. A wider group-type report for facilities managers, with every space inside every category for context. A wider-still department report organised around organisational units instead of physical zones.

Each tier draws from the same Tiger DB SQL and the same pre-calculated metrics, but the prompt, language, and presentation change with the reader. A board member who receives a fifty-page facilities report will not read it; a facilities manager who receives a one-page executive summary cannot act on it.

One design rule sits underneath everything else: the language model never sees raw data. Python computes every utilisation percentage, every action flag, and every department aggregation first. The prompt is a structured payload of pre-verified numbers plus a task description. The model writes narrative around verified arithmetic instead of doing the arithmetic itself, which is what makes the output auditable.

Data sovereignty was a hard constraint. The architecture targets AWS Bedrock in a UK region with Langfuse for observability, and the strategic preference throughout was for private models over public AI APIs so customer occupancy data never crosses jurisdiction.

• Python AI reporting service: data extraction, pandas pre-calculation, model call, structured report write.
• Thin PHP display layer inside the existing SmartScan codebase, with PDF export.
• AWS Bedrock UK region, Azure UK South as a code-free fallback path.
• Langfuse traces for every model call, scheduled-only generation to cap operational cost.
• Model-as-judge validation during development, source-anchored arithmetic at runtime.
• Architecture parameterised on site_id so the pilot extends across the estate without rework.