The Feedback Loop
A Nyckel prediction endpoint is not a fixed artifact. It improves over time as your application provides feedback on its predictions.
How the loop works
1. Your application sends an input
An image, text string, or structured data object is sent to the prediction endpoint via the API.
2. Nyckel returns a prediction
The endpoint returns a label and a confidence score immediately.
3. Your application routes based on confidence
High-confidence predictions are accepted automatically. Low-confidence predictions are sent to a review queue for a human to verify.
4. A human reviews and annotates
The reviewer confirms the prediction is correct, or selects the right label if it was wrong. This creates an annotation.
5. The endpoint retrains
Nyckel automatically incorporates the annotation as a new training sample. The endpoint retrains in the background. No deployment step required.
6. Accuracy improves
Over time, the model sees more examples of the cases it found difficult. Its confidence on similar inputs increases, and fewer predictions require review.
What this means in practice
Early in deployment, you may review 30–40% of predictions. As the endpoint learns from feedback, that number typically drops to under 5% for well-defined classification tasks.
This also means you do not need a large, perfectly labeled dataset before you can deploy. You can start with a small set of examples, go live, and let real-world data improve the endpoint over time.
Building review into your application
To get the most out of the feedback loop, your application should:
- Store predictions — keep a record of every input and its prediction result
- Route by confidence — send low-confidence predictions to a review interface
- Submit annotations — use the Nyckel API to submit corrections after review
- Track accuracy — monitor how confidence distributions change over time
Related
- Confidence scores — understanding and using the confidence value
- Build a feedback loop — implementation guide with API examples