I documented the high‑level architecture, use‑case, class, and sequence diagrams for the Hiver AI Support System, which processes a Customer Tweet through normalization, intent classification, safety gating, triage, retrieval, drafting, and dispatch or escalation.
- System architecture flow: Customer Tweet → Text Normalizer and PII Sanitizer → Intent Classifier → Safety Gate → (Hazard/PII/Fraud) → Triage Engine OR (Routine Support) → ChromaDB Vector Store → Grounded Reply Drafter → Triage Engine → Escalate to Tier‑2 Human Specialist OR Auto‑Handle via Auto‑Reply Dispatcher.
- Use‑case diagram defines six use cases (UC1–UC6) involving Customer, Tier‑2 Agent, and Evaluator.
- Class diagram includes five classes: TweetInput, IntentResult, TriageDecision, SupportResponse, SupportPipeline with their listed attributes.
- Sequence diagram shows step‑by‑step interactions, branching on hazard detection versus safe queries, and final JSON response generation.
I placed the Safety Gate immediately after intent classification to isolate hazardous, PII‑containing, or fraudulent content before it reaches retrieval or drafting components, ensuring safe handling.
The Triage Engine consumes both the draft reply (when safe) and the hazard detection result (when unsafe) to decide whether to Escalate to a human specialist or Auto‑Handle via the dispatcher.