Hey Product Hunt! I'm building GhostSwitch.

The question behind this project is simple: how much more can we get out of a small model before we need a bigger one?

There are already people training specialists (LoRA adapters) for specific tasks and sharing them on Hugging Face. But finding an adapter, checking compatibility, loading it and deciding when to use it still takes work. GhostSwitch brings those pieces together so you can try specialists in an ordinary conversation.

You ask a question. GhostSwitch selects an enabled specialist, attaches it to the model already in memory and generates an answer. If none fits, the base model answers. You can also pick a specialist manually, which is useful when you want to compare results.

The interesting part is whether that specialist actually helps.

An adapter can have a convincing name and model card without producing better answers. That's why GhostSwitch includes evaluations against the plain base model. So far, the coding specialist has shown a measured improvement in the project's evaluation. Most other measured specialists haven't demonstrated an improvement yet. Those that fail the quality gate are excluded from automatic selection.

I want those results to be visible. You can inspect which specialist answered, the route used to select it, generation speed and the checks applied to the response. The measurements and limitations are published in the repo.

This is still an experiment. Routing can choose the wrong specialist. Some adapters won't help your task. A useful result might be discovering where the base model already does the job well enough.

The adapter authors deserve the credit for their training work. GhostSwitch packages and routes their specialists, and gives you a way to explore how they behave together.

Try it: github.com/hiddenway/GhostSwitch

I'd love to hear about a specific task you tried. What did you ask, which specialist did GhostSwitch choose, and was the answer more useful than the base model's?