Intelligence · Baseline from MVP
AI Optimisation Layer
The meta layer. Reads the whole twin, thinks along, writes only recommendations. Vendor-agnostic (vendor-agnostisch), per-tenant budget cap.
What it does
01
Three output stages
Insight (with evidence) → recommendation (with expected benefit) → module proposal (with ROI).
02
Evidence required
Every output carries a reference to real events or twin records. No evidence, no UI output.
03
LLM router
taskType-based: long summaries → Claude, structured extraction → JSON mode. Caching cuts cost.
How it runs
- 01
Event log → feature store
Materialised views on Postgres. The AI never reads the live tables.
- 02
Jobs run
Anomaly + forecast + mining + clustering + summarisation — each on its schedule.
- 03
Recommendation service
Recommendations get `evidence[]` and flow to UI / Slack / email.
What you get
- Insights feed in the dashboard
- Weekly executive report
- Module proposals with ROI estimate
Requires
- → Sites + People + Tasks + Inventory.
In the example
Zero-retention
with LLM vendors contractually
per tenant
budget cap + alert at 80 %