Project case
City incident routing: from manual triage to a five-second recommendation
Based on existing project material. Client, city and selected business fields are anonymized because publication authorization is limited.
Context
A city-operations workflow received about 3,600 incidents per day from hotlines, field patrols, video detection and departmental reporting. Manual reading, classification and assignment created delay and inconsistent decisions.
Delivered approach
Rules, a governed knowledge base and an LLM jointly identify type, location, urgency and responsibility, then provide an explainable routing recommendation. Low-confidence cases go to human review and corrections return to evaluation.
- Incident normalization and address resolution
- RAG citations for responsibility and procedure
- Approximately five-second routing
- Human review, audit and feedback
Recorded outcomes
Existing material records approximately 3,600 incidents per day, around five seconds for intelligent routing, about 30% processing-efficiency improvement and roughly 3× multimodal review throughput. Results depend on data, scope and operating model and should not be assumed for another project.