The author uses generative AI models as opponents to challenge hypotheses and identify vulnerabilities in several AI projects, including passenger flow control, continuous monitoring, and vehicle inspection. Instead of asking for solutions, the author prompts models to find assumptions that would make a solution fail, focusing on what needs to be verified in data, code, and with the client.
In the passenger flow project, the author identified the need to separate video data, telematics, and business logic. A critical issue was discovered regarding time synchronization: screen clocks in videos often had a constant offset from the file timestamps. This led to a structural change where one circuit verifies video and telematics to form candidate intervals, while another handles detection and tracking.
To avoid bias and 'hallucinations' of agreement, the author employs a multiple-model approach. One model may defend a solution, a second identifies failure conditions, and a third compares verifiable claims. This process is designed to ensure that AI suggestions are converted into testable actions rather than accepted as certificates of reliability.
The author emphasizes that AI cannot replace engineering rigor, such as penetration tests, security reviews, and interface contracts. To protect sensitive data, the author uses anonymized or synthetic examples. The final goal is to ensure that the system can report insufficient data or disputed events rather than providing a confident but incorrect number.
