Sort requests.
Teach it which team, topic, or queue a request belongs in. Your app uses the answer to send it to the right place.
> “Which team should handle this?”
System 01
Zils
Teach AI to make decisions your way.
Show Zils examples with the right answers. It trains a small model your app can use to sort requests, flag problems, or score matches.
02How it works
Show it what a good answer looks like.
Show Zils examples of decisions you already know how to make. It trains a model for that job, checks how well it does, and hosts it for your app to use.
For example: which team should handle this customer message?
1Examples
You show it the answers.
Past messages, paired with the right team.
“Please send me a receipt.”
Billing“The app closes every time I open it.”
Technical support“I need to change my account email.”
Account changes
Bring your checked examples in a spreadsheet.
2Training
Zils learns the job.
A small model, trained on your examples. Separate examples test how well it learned. You review the results.
3In your app
A new message comes in.
Your app gets the answer and can route the message.
Message
“My invoice page keeps crashing.”
Example answer
Technical support
Why
It mentions an invoice, but the problem is a broken page.
Illustrated workflow, not live predictions. A few sample rows are shown; useful training needs varied examples for every answer.
03Jobs
Same idea, different jobs.
Choose a team. Flag a problem. Score a match. Pick one repeatable task where you can show what a good answer looks like. Start there, then test how well the model learns it.
Flag problems.
Show it examples that need attention and examples that can wait. Your app can use the result to ask someone to review.
> “Does this need a closer look?”
Score matches.
Provide pairs with scores you have checked. Teach it how well a product, article, or result fits a request.
> “How well does this fit?”
Rank options.
Give Zils 2–10 options and a criterion, such as which support action best addresses a customer’s issue. The SDK returns an ordered list with relative scores.
Optional grouped ranking compares several options together: ten options use 10 judgments instead of 90. Pairwise ranking stays the default.
Fewer comparisons can produce a worse order. Test later choices as well as the first. Relative scores are not confidence percentages.
> “Which action best addresses this issue?”
04What you get
What you get.
You bring the examples. Zils handles training, evaluation, and hosting for accepted models.
A model for your task.
Trained on your examples, with the training run and model version saved in your workspace.
Modelillustrative
namesupport-router
versionv3
training runrun_0418
examples1,024
Results you can check.
See how it performs on examples kept out of training, including the mistakes. Review the results before relying on it.
Held-out examplesillustrative
Billing92%
Account changes88%
Tech support84%
Mistakes to review31
A way to use it in your app.
Once your model passes evaluation, Zils hosts it. Your app sends a request through the API and gets an answer back.
Requestillustrative
# ask your model (illustrative) curl $ZILS_API_URL/v1/systemone \ -H "Authorization: Bearer $ZILS_API_KEY" \ -d '{"model": "support-router", "state": "My invoice page keeps crashing.", "questions": {"decision": "Which team?"}}' # → a decision for your app to use
Before you upload.
Training examples are copied to approved workers, whose operators can read and retain them. This is not confidential computing. Use only data you are authorized to share.
Read how data is handled05Research
See what training changed.
We trained a model to choose the next action in a customer-support conversation, then tested it on conversations it had not trained on.
On 500 test conversations from the public ABCD dataset, our trained model chose the recorded answer 79.2% of the time, compared with 70.6% for TypeSafe JEV, another decision model. That is 29.3% fewer mistakes on the same next-action task.
+8.6 percentage points vs TypeSafe JEV
JEV was tested live; Zils scores reuse verified predictions from the same frozen test set. One research experiment, not a guarantee for your task. It tested the choice of the next action, not success at resolving a whole conversation. JEV was more accurate on its smaller set of high-confidence answers; Zils still overstates its confidence. The research model shown here is not deployed in the shared playground.
06Start
Start with a decision
you know how to make.
Bring examples and their right answers. The training workspace helps you prepare your data and start a run.