Next Steps

You’ve created a classification function, defined labels, invoked predictions, and set up a review workflow that trains a private model on your data. That’s the full happy path. From here, the work is mostly about deepening — using confidence well, wiring feedback in from your application, and building the function into a real production integration.

Use confidence well

The single highest-leverage change you can make once the function is live is using the confidence score deliberately:

• Pick a high-confidence threshold above which you act on predictions automatically.
• Pick a low-confidence threshold below which you route to a fallback or to a human.
• Re-check those thresholds as your model improves — what was a sensible cutoff at 50 samples is often too conservative at 5,000.

The full guide: Tuning Confidence Thresholds.

Build a feedback loop from your application

The review queue is one source of training data. The other — usually larger over time — is feedback from your own application. Whenever a user, support agent, or downstream process disagrees with a prediction, pipe that signal into Nyckel via the annotation API.

The full guide: Build a Feedback Loop.

Plan for production

A few things worth thinking about before you ship:

• Latency. Nyckel endpoints typically respond in under a second, but plan for variability. Don’t put a synchronous Nyckel call on a critical path without a timeout.
• Retries. Treat invokes as you would any external HTTP call. Retry transient failures; don’t retry 4xx responses.
• Cost shape. You pay per invoke, not per train. High-traffic functions add up — consider caching predictions for identical inputs or sampling traffic for review rather than reviewing 100%.

The full guide: Production Integration Patterns.

Improve accuracy over time

The function is never “done.” Accuracy improvements come from:

• Reviewing predictions on a steady cadence (not in bursts)
• Prioritizing low-confidence and corrected predictions
• Adding samples for edge cases the model hasn’t seen
• Sometimes refining the labels themselves as you learn what they should mean

The full guide: Improve accuracy over time.

Go deeper on the platform

• Developer Platform — architecture, principles, and reference for building production integrations.
• Prediction Endpoints — full reference for the invoke endpoint and its options.
• API Reference — complete REST API documentation.

Or try another function type

• Create Your Own Box Detect Function — find and locate a specific object in images.
• Create Your Own Search Function — query a reference corpus by meaning.