Principles

Nyckel is built on three architectural beliefs about how ML should work in production.

Inference and training belong in the same system

Most ML platforms treat inference and training as separate concerns — you deploy a model, and improving it is a different system, a different team, a different workflow. Nyckel is built around the opposite belief: invoke and annotate should be two operations on the same API, so that every prediction your application makes is a potential training sample and every correction feeds directly back into the model. Keeping the inference stream and the training pipeline unified isn’t a convenience — it’s what makes a production endpoint that actually improves over time.

Start fast, improve as you go

Getting started with ML should take minutes, not months. You shouldn’t need to select a model architecture, configure training infrastructure, or define an evaluation pipeline before you can test a classification idea. Start with a working endpoint, get real predictions flowing, then use feedback to improve accuracy. The right time to tune is after you’ve seen production data — not before.

Engineers and operations teams need to be in the same loop

Engineers write the invoke call; domain experts and operators review predictions, correct mistakes, and label edge cases. When those two groups work in separate tools, corrections never reach the model. Nyckel makes annotation a first-class API operation so every correction goes directly to the training pipeline. And because every invoke response carries a confidence score, your application can decide programmatically when to act on a prediction and when to escalate it for human review — with the full annotated history available for testing and auditing.