What Makes a Good Applied ML Project

A good applied ML project has a difficult model problem, but it also has a clear path from input to decision.

A concrete unit of work

Define what the system receives, what it returns, and who uses the result. “Analyze technique” is vague. “Estimate stroke phases and compare peak trunk timing across sessions” is testable.

A baseline that can lose

Start with a rule, linear model, or simple temporal baseline. A complex model should earn its place through a measured improvement, not through architectural novelty.

Failure analysis

Aggregate accuracy hides where the system breaks. Split results by capture condition, athlete, motion phase, and input quality. The useful outcome is often a map of limitations rather than one score.

An operational boundary

The model is one stage in a system. Data capture, preprocessing, latency, correction workflows, and output presentation determine whether the work can be used.