Subject alignment
A gate that knows what your system is for.
A small model that decides whether a message, document or request is about your subject before anything else runs. Off-topic and out-of-scope inputs are refused in milliseconds, so your other models and your people only see what belongs to them.
Incoming
→ Support bot
✕ Refused before the bot runs
What it does
Built for tasks like these.
- Refuse off-topic chat in a support bot
- Screen uploads before a pipeline runs
- Keep an agent inside its job
- Filter a feed to one subject
How it works
From your examples to a model you own.
- 01
Show it your scope
Pull a few hundred real requests and mark which belong to your system and which don't, especially the close calls.
- 02
Train the gate
We train a small adapter on those labels, on a laptop, in an afternoon.
- 03
Put it in front
Every request passes the gate first. Out-of-scope ones get a polite refusal before anything expensive runs.
Private by default
Your requests never leave the building.
Every input passes through the gate, so the gate sees everything. Running it on your own hardware means the screening step never becomes a data leak of its own.
- Your customers' data leaves the building on every call
- The model changes whenever they ship
- You pay for a huge general model to do one narrow task
- The request never leaves the device
- The model is a file you own; it changes when you retrain it
- Small enough to run on hardware you already have
Trained, not prompted
Accurate on your task.
Trained on examples of what is and isn't in scope for you, not a generic moderation policy. It learns your edges: the near-miss questions a prompt never anticipates.
InputCan you help me write a cover letter? (sent to a tax-prep support bot)
General model, prompted
Of course! Here's a draft cover letter tailored to…
Micromodel, trained on your examples
out_of_scope · "I can help with your return and documents. For anything else, try a general assistant."
Illustrative example.
And it costs less
No per-call bill.
It runs on every request, which is exactly where per-call pricing hurts. A local gate costs nothing per call and turns away the traffic you would otherwise pay a large model to answer.
$120,000 annual savings
See it for yourself
Watch a micromodel run on your own GPU.
A guided demo, about a minute. It downloads once and runs entirely in your browser; nothing you type is sent anywhere.Open the demoSort and judge
Related solutions.
Have a task like this?
We'll tell you if a micromodel can do it. Or book a demo, or email hello@micromodel.company.