customer-support-classifier
A production-ready fine-tune recipe for routing customer-support inbox traffic. Outputs a priority tier, a topic label, and a suggested next action — trained on a synthetic, de-identified ticket corpus and benchmarked against a held-out gold set.
Install
$joseki install customer-support-classifier copy
Overview
Overview
This package fine-tunes a base model to triage customer-support tickets. It outputs a structured decision — priority tier, topic label, and a suggested next action — trained on a synthetic, de-identified ticket corpus and benchmarked against a held-out gold set.
Inputs
- Ticket subject
- Ticket body (free text, any language supported by the base model)
- Optional account metadata (tier, region)
Outputs
- Priority: `urgent` | `high` | `normal` | `low`
- Topic label (one of the configured queues)
- Suggested next action (one of: respond, escalate, request-info, close)
When to use
Use this package as a first-pass classifier in front of a human queue, or as the routing step in an automated pipeline.
Eval results
label_accuracy
format_following
pii_leakage
tone_consistency
Safety scan
Package signing
Signed by Northwind AI
Prompt injection scan
No vulnerable templates
License audit
private-commercial
Artifact hashes
sha256 9f1c5b3e2a8d4c7e1f0b9a3d4e5c6f7a8b9c0d1e2f3a4b5c6d7e8f9a0b1c2d3e manifest.yaml sha256 1a2b3c4d5e6f708192a3b4c5d6e7f80910abcdef0123456789abcdef01234567 training-data.tar.zst