CannAI Solutions builds local, auditable AI for plant health in controlled environment horticulture. I founded the company and built the stack myself: eleven models, among them a production 4-class classifier covering pest, disease, nutrient and healthy, trained on 259,000 labelled images. The problem I work on is not accuracy. One symptom has several possible causes, so a share of errors are not near misses but reversals: feed more instead of less, treat instead of correct. Nobody reports that error class, not in benchmarks and not in acceptance tests. We name it, measure it, and control it with directed rules and abstention calibrated against the cost of being wrong. About 300 training runs since 2025. In September 2026 the RWTH Aachen IT Center granted 10,000 GPU hours.