# Inference savings

The business model is the one an energy service company uses: a small custom model is trained for each repeatable action on the customer's own server, and HELM is paid a share of the savings. The first step is distillscan, an open tool that works. It is an offline scanner over exported LLM traces that clusters the workloads and ranks them by estimated annual savings, so you see what is ready to move to a cheaper model.

Idea. A service that watches real AI traffic and moves expensive model calls to small specialised models, paid from verified savings.

Status: drawer
