It’s great that AI models can code, draw images, and recognize faces.
We have become accustomed to working with a very powerful beta system: we use what we like and we ditch what we don’t.
In production, AI looks very different. Reference standards. Benchmarked outcomes. Real-world implementation. Known margins of error.
Bringing an AI model to a production standard requires procurement of curated, diverse, cleaned, standardized real-world data, to a defined scope, and a traceable development trajectory. It needs testing in real and simulated settings to measure its performance to calibrate what its predictions are worth.
There are not many of those examples yet. Performance analysis is expensive. Still, performance analysis may be the only way forward for AI. Settling for eternal beta testing will not lead to structural improvements in areas where lives are at stake. Such as in TB diagnostics.
Proud to have contributed to the design and implementation of a sandbox Validation Platform for AI-driven TB diagnostics. Its latest achievements have been published in the “Evaluation of CAD products in community and facility TB screening: 2025 policy update ” paper in ERJ Open Research.