Research

Research is TERRAIN’s “fail fast, learn faster” engine. Instead of betting months on one approach, the team builds Minimum Viable Products — deliberately basic versions of the AI model — to find out quickly what works, what doesn’t, and what users actually need.

What happens in this phase

• Build MVPs, not monuments. Basic model versions explore different techniques and functionalities at low cost. Pre-trained models and transfer learning speed things up wherever they fit.
• Experiment systematically. Different algorithms, hyperparameters, and feature-engineering approaches are tried, and every experiment’s results are recorded for comparison — this is where the model registry habit begins.
• Get user feedback immediately. Feedback methods are wired in from the first prototype, so real users shape direction while changing course is still cheap.
• Kill weak options early. The point of the phase is elimination: promising paths earn investment, dead ends are documented and closed.

Watch out for

• Prototypes that quietly become production systems without ever passing through rigorous training and evaluation.
• Experiments without records. An untracked experiment teaches one person once; a registered one teaches the organization forever.
• “One more iteration” syndrome — the phase needs explicit criteria for when an MVP has answered its question.

Method spotlight

Research runs on MVP Experimentation (the Phase-R hexagon cycle: data preparation → rapid prototyping → pre-trained models → tuning → justified decisions → deployed MVP) and Lean Innovation. Diagrams on the framework page.

Go deeper

The prototyping and backlog chapters (10, 14–16) cover the delivery mechanics — free here from September 2026, or in the full edition on Amazon today.