Building Better ML Projects from the Ground Up

The problem is that projects fail slowly - after months of investment, unclear progress, and data scientists feeling guilty for things beyond their control.

Many failure modes are predictable and discoverable early: weak signal, unstable data, poorly defined objectives, or integration constraints. Unfortunately, most ML initiatives are not designed to detect these risks until late in development due to optimism bias and bad framing.

This talk reframes ML development through the lens of experimental design.

We will cover:

- Why ML project failure mirrors the scientific replication crisis (metric fishing, test-set reuse, and selective reporting)
- Applying Design of Experiments (DoE) thinking to ML project planning
- Defining success criteria and kill conditions before training the first model
- Staged investment frameworks: the minimum experiment needed to justify continued development
- Reframing failed projects into successful learning

The goal is credible successes and inexpensive failures that accelerate learning and improve organizational trust in ML outcomes.

Attendees will leave with a practical framework for structuring ML initiatives that produce decisive answers instead of prolonged tragedies.