Improve Accuracy
A Nyckel function gets better the more it’s used and reviewed. These guides cover the practical “how” of making that happen on a steady cadence rather than in bursts.
The accuracy guides
- Review and improve predictions — how to build a sustainable review workflow that turns production traffic into training data.
- Get reliable predictions — designing labels and preparing data so the function’s predictions are trustworthy from the start.
- Improve accuracy over time — using feedback, targeted samples, and annotation strategy to drive long-term accuracy gains.
- Tuning Confidence Thresholds — picking the cutoffs in your application that decide what to act on automatically vs. what to hold for review.
The two highest-leverage actions
If you only have time for two things:
- Review predictions on a steady cadence, prioritizing low-confidence ones. Each review is a training sample, and low-confidence reviews move the model the most.
- Use confidence thresholds deliberately in your application — and re-check them as your model improves.
Everything else is a refinement on those two.
Next
- Build Production Workflows — patterns for using Nyckel in production.
- Explore the API Reference — the full REST API.