George Firican, founder of Lights On Data, sat down with Ho Bae to talk about why AI that works in a demo struggles once it meets real company data.
The conversation starts from one question: if the data is AI-ready, ready for what? The same dataset can work for one AI job and fail for another. They walk through real examples, from service case follow-up decisions to churn prediction and a food supply quoting project.
What they talked about
- 01What AI-ready means, and why it depends on the task
- 02Filling missing cases with DTS, and why naive synthetic data fails
- 03Swapping sensitive values for context-preserving stand-ins with LLM Capsule before they reach AI
- 04Checking whether prepared data actually helps the task, and approval criteria over metrics
- 05Build vs. buy, and reusing what still applies
