电商客服意图路由沙盘 (E-commerce CS Routing Sandbox)

An e-commerce customer-service quality demo. It simulates dialogues and shows, message by message, how Jev judges intent, emotion, urgency and risk, how a decision table routes the conversation to a skill group with a queue, priority and SLA, and how dangerous agent phrasing is blocked before sending.

For user messages: first-level intent (14 classes), emotion intensity (5 levels), urgency (4 levels), whether to escalate to a human, and user-side risk type such as regulatory complaint, media exposure, litigation, personal attack or suspected abuse. Agent messages get their own set: dangerous phrasing type, service quality, empathy and actionable plan.

No — without a key the app runs a built-in Mock rule engine whose output has exactly the same structure as Jev, and 12 built-in scripts replay with one click. For real judgments, copy .env.example to .env, add a TYPESAFE_API_KEY from console.typesafe.ai, and the engine badge switches from Mock to Jev jev-latest.

Per the README: the API bills $0.042 per million input tokens and output is not billed for now; one message is about 200–400 tokens, so a 10-turn session comes to under $0.0002. The dashboard shows the current session's request count and estimated cost in real time.

An ordered decision table where the first match wins: conditions test choice membership, score level indexes or noul probabilities, with all conditions ANDed inside a rule and ORed inside any. Eight rules send risk, complaint, anger, logistics and escalation cases to the right queues with SLAs; the README explains why R3 and R6 precede R4.

It checks agent messages before they are sent. Block-level phrasing — unauthorized compensation, privacy leaks, emotional breakdown, private trading, illegal claims — is stopped; warn-level phrasing such as absolute promises, hard time commitments, blame-shifting and disparagement gets a rewrite suggestion. The probability threshold defaults to 0.5.

Almost everything is data: shared/policies.json holds the judgment items, criteria and emotion levels, routing.json holds the routing rules, queues, SLAs and intercept levels, and scripts.json holds the demo scripts. The server re-reads them per request; npm test runs 34 unit tests and npm run typecheck type-checks the code.