Nyckel's Dataset-First ML
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
Nyckel is trying to turn machine learning from an expert workflow into a simple product workflow. Instead of asking a customer to pick a model, tune settings, and read model metrics, it asks for labeled examples, trains several candidate systems in parallel, and shows whether the output looks right on the customer’s own data. That makes ML feel less like buying tools for engineers and more like calling an API for a specific job.
-
The key product choice is that the abstraction is the dataset, not the model. Customers upload text, images, or tabular rows, label examples, and Nyckel automatically retrains after data changes and estimates accuracy with cross validation, so the loop is centered on examples and predictions instead of model management.
-
This puts Nyckel closer to an end to end ML utility than a classic MLOps tool. In the interview, the stack combines in house and open source models inside a parallelized AutoML engine. That is similar in direction to Vertex AI, but Nyckel is aimed more squarely at product managers and developers who want a working classifier fast.
-
The deeper business implication is that model providers can become interchangeable suppliers. If the customer only cares whether their invoice scanner, moderation filter, or image classifier works on their own examples, the durable value shifts to data ingestion, fast evaluation, automatic retraining, and deployment plumbing, not to exposing a branded model choice.
This market is heading toward bundled products where labeling, training, testing, and deployment collapse into one workflow. As foundation models improve, the winners are likely to be the companies that hide more of the stack, shorten time to a working result, and earn trust by showing concrete performance on a customer’s own data within minutes.