Health technology startup · Prior founder experience
Removing the data bottleneck holding back ML development
Redesigned the data-collection process to produce 10× more usable ML training data in approximately three months. A separate workflow initiative shortened development time by 75%.
The constraint was upstream of the model
The company developed products combining physiological sensors, machine learning and mobile software. Model development depended on collecting usable observations with appropriate reference information.
The bottleneck extended across participant recruitment, onboarding, meal recording and data-quality feedback. Improving the model alone would not repair a process that failed to produce enough usable training records.
Redesign the collection process
Changed how participants entered the research process, recorded meals and received immediate feedback about data quality. The work connected the participant experience with the requirements of the training dataset.
Working with the CTO, engineers and external subject experts, the founder treated data collection as a product and operating problem. The goal was more usable evidence for development, rather than a larger pile of raw sensor readings.
Measured result
The redesigned process produced 10× more usable training data in approximately three months. This is a data-throughput result. It is not a claim of 10× better model accuracy or 10× more customers.
Separately, workflow redesign and automation across marketing, testing/deployment and support shortened development time by 75%. The two results came from distinct interventions and should not be interpreted as a single causal chain.
A second lesson: interpreting the output
Another product iteration moved stress/performance feedback from manual next-day conversations to nightly reports and then real-time graphs. More immediate data did not automatically make the product more useful: users struggled to interpret the graphs.
Threshold-based guidance made the output more actionable. This work illustrates why model and sensor capability must be connected to a decision the user understands. No numerical retention lift is attributed to that specific change.
Why the case matters
The work demonstrates diagnosis of the constraint around an ML system, followed by changes to the operating process that supplied it. Technical progress depended on recruitment, participant behavior, reference records and feedback working together.
Prior founder experience, before Measured Intuition. The historical results describe this company’s work, not a current consulting-client deployment.
Discuss a workflow worth changing.
Start with the operating constraint, the current process and the business result that would justify changing it.