Ecom Cs Agent

It is an e-commerce after-sales customer-service agent built on the learn-claude-code skeleton. Each dialogue turn first gets a structured Jev judgment, and a code-level gate decides whether the main DeepSeek agent handles the message or a human does.

In one call Jev answers two questions in parallel: a choice question for the customer's primary after-sales intent, and an additional needs_human Noul question. The intent answer carries a confidence computed with Jev's (p_max − 1/n)/(1 − 1/n) formula.

When confidence is below the threshold (raised by 0.1 for refund, exchange and address-change intents), when the intent is request_human or complaint_escalation, or when needs_human is at least 0.5. Judge errors such as network, rate-limit or format failures also go to a human, fail-safe.

The author's v1 experiment found intent confidence measures how certain the intent is, not whether the bot can handle it. A customer saying a pair of shoes faded after one wash has a clear exchange intent, but whether that counts as a quality issue needs human judgment.

It ships 30 annotated conversations and compares four arms under the same gate: no judgment layer, the Jev judgment layer, LLM self-reported confidence, and a three-way weighted fusion of self-score, character n-gram similarity and keywords. There is also an LLM-as-judge quality evaluation scoring relevance, accuracy, completeness and usefulness.

The main loop with an eight-turn cap, tool calling for orders, logistics, policy RAG and preferences, three permission gates for refunds, exchanges and address changes, an after-sales subagent, context compaction into .transcripts/, customer memory and four handling skills.

The README visible in the digest is truncated in the skills section and the repository lists no license, so setup instructions and the full comparison results are not documented here. It calls Jev over direct HTTP instead of the official Python SDK because the SDK requires Python 3.10, and offline use needs the MockJudge.