Data Engine Above Models for Classification
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
The durable value in classification sits above the model, in the system that turns messy company specific examples into a repeatable decision service. For a classifier, the hard part is not generating one plausible answer. It is showing the product owner a batch of their own examples, measuring accuracy on that data, spotting failures, relabeling edge cases, and redeploying fast. That workflow is the data engine.
-
Nyckel is built around that loop. Customers upload about 100 labeled text or image samples, often annotated directly in the UI, see cross validated results on their own examples in seconds, then tweak labels or add more data until the predictions match their use case.
-
This is why classification product layers survive even when foundation models improve. The underlying model can be swapped, multiplexed, or outsourced, but the winning product still needs evaluation, versioning, monitoring, and a simple input to output interface for non experts.
-
The market has been moving in this direction. Scale expanded from raw labeling into RLHF and broader model tooling, while Dataiku bundled ingest, prep, AutoML, and visualization into one GUI so business teams can build and check models without stitching together many point tools.
Going forward, more AI products will look less like a single magical model call and more like packaged decision systems wrapped around data feedback loops. As foundation models commoditize, the companies that win classification will be the ones that make it easiest to define the task, verify performance on live customer data, and improve continuously with very little human effort.