How Deepcell reduced its model training expenses by 50% in 2 months

The challenge

More clients meant more data, more requests, and more pressure on the ML workflow. The binding constraint was labeled data: manual labeling was slow, expensive, and could not keep pace with the scale the platform needed. The team brought in outside expertise to rethink the training system itself.

The work

• A self-supervised pre-training system. Built on self-supervised representation learning, adversarial learning, recalibration, aleatoric and epistemic uncertainty, and neural-network architecture design. It reduced model errors by more than 50% and improved embedding quality.
• 90% less labeled data. The new system cut the labeled-data requirement by over 90%, which in turn cut model training costs by more than half and let Deepcell process and sort cells at a scale manual labeling could never support.
• Unsupervised clustering with biologically interpretable outcomes. Researchers could discern patterns and categorize cells without explicit labels, opening analysis that labeled pipelines could not reach.
• R&D roadmapping. Beyond the build, Eventum’s engineers helped chart the technical roadmap: areas of focus, risks, and sequencing aligned to Deepcell’s goals.
• Mentorship and hiring support. Mentorship for Deepcell’s ML scientists and engineers, plus technical interviewing support for leadership candidates.

The result

Model training costs down by more than 50% within two months. Labeled-data requirements down by over 90%. Model errors down by more than half. And a training system that scales with data volume instead of with labeling headcount.

Why it mattered

For an AI company whose product is analysis at scale, the cost curve of training is the business model. Self-supervised pre-training changed Deepcell’s curve: less labeling, cheaper training, better embeddings, and new kinds of analysis, all from the same data they already had.

Conclusion

Three Eventum consultants helped Deepcell replace a labeling-bound training pipeline with a self-supervised system: 50% lower training costs, 90% less labeled data, and measurably better models, delivered in two months.