Bundling ML Stack into One Product
Oscar Beijbom, co-founder and CTO of Nyckel, on the opportunites in the AI/ML tooling market
The real advantage of bundling the ML stack is that it turns machine learning from a specialist workflow into a product workflow. When data labeling, training, error analysis, deployment, and monitoring live in one place, the customer can upload examples, see model mistakes, fix labels, retrain, and ship without stitching together separate tools. That matters most for smaller teams and domain experts, because the hard part is rarely one model, it is the loop of improving a model on their own data.
-
The practical synergy is in the handoff between steps. Labelbox now groups training data, model runs, error analysis, versioning, and one click training inside experiments, instead of treating labeling as a separate system. That makes the output of one step immediately usable by the next one.
-
The integrated product also changes who can operate ML. Nyckel describes the winning interface as inputs and outputs at the data layer, where a product manager or developer can label 100 examples, validate results with cross validation, and deploy quickly without managing models directly.
-
The clearest large company analogue is Vertex AI. Google bundles training, pipelines, registry, serving, evaluation, and monitoring into one platform. That shows where the market is heading, toward fewer point tools and more systems that manage the full model lifecycle in one workflow.
The next phase of the market is less about selling isolated labeling seats and more about owning the feedback loop around customer data. The winners will be the companies that make retraining and improvement feel continuous, where every mistake found in production flows back into labeling, evaluation, and the next model version inside the same product.