Challenge
A shared administrative team receives finance, student-record, digital-learning and sensitive-support enquiries. The system must identify intent, detect missing information and prevent sensitive cases from reaching answer generation.
Selected work
Azure AI engineering · staff triage · 2026
Agentic Support Intelligence
A staff-facing university enquiry workbench that turns unstructured requests into a safe route, clarification request or cited draft response.
My role
I designed the ingestion, Azure hybrid retrieval, structured classifier, deterministic routing policy, evaluation harness, FastAPI boundary and staff review interface.
Outcome
The working system indexes eight approved sources across four answerable domains. Either conservative rules or the classifier can escalate a case, while code prevents every escalated case from entering RAG generation.
At a glance
4answerable domains
21evaluated triage cases
0sensitive generation leaks
Staff enquiry workbench
From unstructured enquiry to a safe next action.
Paste a synthetic enquiry. The system classifies it, applies deterministic routing, and only retrieves evidence when generation is allowed.
Public demo mode
Open staff queueResults are not stored. Use synthetic enquiries only.
Awaiting enquiry
The structured classification, routing decision and cited draft will appear here.
System view
The workflow.
A simplified view of the stages and boundaries that shape the project.
- 01Staff enquiry
- 02Safety rules
- 03Structured classification
- 04Routing policy
- 05Category-filtered RAG
- 06Human review
Approach
Decisions that shaped the work.
Make escalation one-way
Rules and model classification can both escalate an enquiry. Neither can override a sensitive referral and send it to generation.
Route before retrieval
Only answerable categories reach Azure hybrid search; the selected category becomes an enforced index filter rather than a user-controlled document choice.
Keep staff accountable
Generated text is a draft with source citations. Persistence records the action and review state, while sensitive raw text is redacted.
Findings
What the evidence says.
- The corpus separates corporate finance from operational expense guidance, avoiding retrieval from the wrong financial source.
- Minerva enquiries can identify missing module, assignment and error details before staff draft a response.
- Sensitive cases return a specialist route without retrieval, citations or a generated substantive answer.
Engineering reflection
This is a decision-support prototype, not an autonomous case-management system. Production use would require institutional policy review, Entra-authenticated staff access, independent safety evaluation and agreed retention controls.