Build Production Workflows

A Nyckel function is a live endpoint, which means putting it in production is mostly about treating it like any other dependency — with timeouts, retries, and a plan for what to do when something goes wrong — plus a few patterns specific to ML predictions.

The developer platform guides

• Developer Platform overview — the architecture and what’s available to you as a developer.
• Principles — the design principles behind the API that make integration predictable.
• Prediction Endpoints — full reference on the invoke endpoint, options, and behavior.
• Build a Feedback Loop — capturing user/system signals from your application and turning them into Nyckel annotations.
• Production Integration Patterns — practical patterns: synchronous vs. asynchronous, low- vs. high-confidence routing, caching, retries.

Three patterns worth knowing

• Confidence-routed actions. Act on high-confidence predictions automatically; route mid-confidence to a fallback (older logic, simpler rule); send low-confidence to human review. The thresholds are the only part you have to tune.
• Feedback loop. Whenever a downstream system or user disagrees with a prediction (refunds reversed, support tickets re-routed, content un-flagged), pipe that signal back as an annotation. Production usage becomes training data automatically.
• Sample-based review. You don’t have to review 100% of predictions. Reviewing a representative sample at a steady cadence often produces better accuracy gains than reviewing exhaustively for a week and then nothing for a month.

Next

• Learn Core Concepts — the conceptual foundation.
• Explore the API Reference — the full REST API.