Helping engineers detect edge-case safety issues

Vision-based AI is already elevating road and driver safety, and over the next 2-5 years, I see it moving from passive monitoring to an active safety partner that helps fleets prevent incidents. Instead of just recording events, AI will increasingly predict risk by analysing real-time driver behaviour, vehicle health, and environmental data, and then intervene before an accident occurs. I hope that this shift drives a drastic reduction in preventable accidents by acting as an always-on coach that doesn’t just flag errors but also actively predicts behaviours of road users and advances safety for all.

To achieve this, I expect AI to help engineers detect edge-case safety issues using large-scale simulation data, analyse millions of miles of driving logs for anomalies, and recommend design improvements based on pattern recognition. AI can also help explain why a model might fail under certain road conditions, automatically generate test scenarios based on identified weaknesses, and act as an engineering “second brain” for functional safety reviews.

To truly revolutionise fleet safety and management, we need AI to automate the heavy lifting of data interpretation. Semantic Video Search should handle complex, natural language queries like “Show me all instances of distracted driving near school zones in rainy conditions,” and instantly retrieve relevant footage from the edge without manual sifting. Predictive Risk Modeling should also take over the continuous monitoring of risky behaviours and identify near-miss clusters and high-risk behaviours that haven’t resulted in accidents yet, allowing fleets to proactively alter routes or training.

Automated Coaching should provide the routine nudge with realtime, positive reinforcement that scales better than human intervention ever could. When multiple risks appear simultaneously, for example, heavy traffic combined with poor weather, then AI should always prioritise threats, give meaningful warnings, and reduce alert fatigue by filtering noise.

Build for the messy, unpredictable, high-stakes physical world

Looking further ahead, I anticipate AI to possess intent reasoning. It’s not enough to detect a pedestrian, the system must also predict if that pedestrian is about to step into traffic based on subtle cues. I envision persistent world models at the edge, where a single pass through an intersection updates a dynamic map of risk, traffic patterns, and infrastructure health, serving multiple business goals simultaneously without needing new sensors. For AI startups, the challenge is clear : Don’t just build for the data centre. Build for the messy, unpredictable, high-stakes physical world. That is where the next generation of value and safety will be created.