Computer vision in the enterprise world has a branding problem. Everyone immediately thinks of factory floor defect detection — cameras pointing at a production line, looking for scratches or misalignments. That is a valid application, and we have done it. We built a system for a manufacturing client that outperformed the inspection solution from a competitor with a market cap north of $40 billion. But defect detection is maybe 20% of what computer vision can do for organizations.
Our most rewarding computer vision project was in conservation. Working with environmental researchers, we deployed camera-based monitoring systems to track wildlife in their natural habitat. The system identifies species, counts individuals, and detects behavioral patterns that would take human researchers thousands of hours to observe. A camera that runs 24/7 for a year generates more observation data than a team of researchers could collect in a decade. The AI processes it all.
In education, we built a computer vision system that won a national award. The application was not what you might expect. It was not surveillance or exam proctoring. It was helping students with learning differences by analyzing their engagement patterns during lessons. The system could identify when students were struggling or disengaged without singling them out. Teachers got aggregate insights about lesson effectiveness. No individual student was tracked or scored. It was computer vision in service of better teaching, not better monitoring.
Retail is another area where computer vision is underused. Most retailers have security cameras covering every square meter of their stores, and the footage is only reviewed after something goes wrong. We worked with a retail group to analyze this existing footage for customer behavior insights: traffic patterns, dwell time at displays, queue lengths. No new hardware was needed — just AI models applied to cameras that were already there. The store redesign based on this data improved sales per square meter measurably.
The manufacturing project I mentioned is worth describing because it illustrates a common pattern: incumbents with expensive solutions that are worse than what a small team can build with modern tools. The existing inspection system from the large vendor required special hardware, special lighting, and a team of three to maintain. It caught about 85% of defects. Our system used standard industrial cameras, required no special lighting adjustments, ran on commodity hardware, and caught 93% of defects. It also cost a fraction of the incumbent solution.
The technical lesson across all these projects: the model is rarely the bottleneck. The bottleneck is data labeling, edge case handling, and deployment infrastructure. We typically spend two weeks on the model and two months on everything around it. Collecting representative training data, labeling it accurately, handling lighting variations and camera angles, building the pipeline that moves images from camera to model to output — that is where the real work lives.
One pattern I find interesting is what happens when you give domain experts access to computer vision tools. Conservation biologists come up with applications that no AI engineer would think of. Retail managers see patterns in customer behavior data that are invisible to data scientists. Education researchers ask questions about engagement that nobody in the technology world has considered. The best computer vision projects come from domain experts who understand the problem, paired with engineers who understand the technology.
Edge deployment is increasingly important for our computer vision work. Many of our applications run on-site, on dedicated hardware, with no cloud dependency. The conservation cameras are in locations with no internet connection. The retail analytics need to process in real-time without latency. The manufacturing inspection needs to run at production line speed. We have gotten good at optimizing models for edge hardware — quantizing, pruning, and distilling until they run on devices that cost a few hundred dollars.
The privacy dimension of computer vision is something I think about a lot. Every camera-based system we build includes privacy by design. In retail, we do not identify individuals — we track anonymous movement patterns. In education, no individual student data is captured or stored. In conservation, there are no privacy concerns because the subjects are animals. But the technology is the same technology that could be used for surveillance, and we are deliberate about drawing a clear line between analytics and monitoring.
Computer vision is going to be everywhere within five years. Not because the technology is new — it is not — but because the cost has dropped to the point where the ROI makes sense for applications that were previously too niche or too expensive. The organizations that benefit most will be the ones that look at their existing camera infrastructure and ask: what else could we learn from this?