Structured decisions trained on your data.

Train a custom classifier in minutes, with confidence calibrated to your data. No ML pipeline, labeling tools, or infrastructure to manage.

Start building

Your decisions should
fit your data.

General-purpose models, like LLMs and decision APIs, make it easy to get structured answers. But their accuracy on your task is unknown until you test them, and their confidence scores may not reflect your data.

Training on your own examples teaches a model the distinctions that matter to your business.

Image

leaf spot0.94
healthy0.04
rust0.02

Text

"I was charged twice for my October invoice."

billing team0.88
support0.09
sales0.03

Data row

amount: $2,847
account_age: 2 days
device_match: false

review0.91
approve0.07
decline0.02

Training your own model shouldn't be so hard.

Nyckel handles model training, deployment, scaling, and labeling tools. You bring the examples.

01
Create a function.

Name the decision you want made.
```
>>> func = (
...   ImageClassificationFunction.create(
...     "LeafHealth", credentials))
```
Functions/New Function

InputImageText

OutputClassifyTag

NameLeafHealth Create Function
02
Add examples.

Label them in Nyckel's built-in interface, or send them through the API.
```
>>> func.create_samples([
...     ("leaf-01.jpg", "healthy"),
...     ("leaf-02.jpg", "leaf spot")
... ])
```
LeafHealth/Train

leaf-02.jpg
healthyleaf spotrust
03
Call the API.

Get a label and confidence score from a live, scalable endpoint.
```
>>> result = func.invoke([
...     "leaf-0412.jpg"
... ])[0]
>>> result.label_name, result.confidence
('leaf spot', 0.94)
```
LeafHealth/Try it

leaf-0412.jpgInvoke

Model predicts leaf spot0.94 confidence Is this right? YesNo

Don't have examples yet?
Start with a ready-made classifier.

Your model should
keep getting better.

When your requirements change or you find mistakes, correct examples in Nyckel. It retrains in the background and switches over only when the new model is better, without rebuilding your pipeline or changing your API integration.

Active learningNyckel picks the uncertain few, a random sample, and rare labels for review. You confirm or correct them.