Support ticket triage and routing
Every incoming ticket tagged, prioritised and routed with your own categories, in milliseconds, on a model small enough to run anywhere.
The problem
Your categories are not the internet's categories. A general model paraphrases instead of picking from your list, and a frontier API is slow and costly for something that runs on every message.
Data you already have
Your ticket history: the message, and the queue, tag and priority a person gave it. Most help desks export this in an afternoon.
What we train
A 0.6-billion-parameter model (or 4B for longer tickets) fine-tuned to answer with exactly your label vocabulary. On a public intent dataset this size of model went from near zero to on-par with a frontier model once trained on the task, at a small fraction of the cost per ticket.
How we measure it
Accuracy per queue and per priority on held-out tickets, against the base model and, if you want, against the frontier API you use today.
Where it runs
Inside your help desk pipeline on a single small GPU, or on AWS Bedrock as an imported custom model. For enterprises with many brands or regions: one base, one adapter per tenant, on one deployment.
- Who it is for
- Support operations, IT service desks, shared-service centres, SaaS companies.
- Typical timeline
- Two to four weeks from intake to handover.
- You receive
- Merged weights, adapter, quantized copy, eval report, deploy recipe, training manifest.