Building a car recognition application (pt. 1) (blog.wildedge.dev)

A prototype for a car recognition application, named CarScanner, has been developed to leverage multimodal large language models (LLMs) like Gemini for real-time vehicle identification. Unlike traditional methods that require extensive image labeling and model training, developers can now point a smartphone at a car and receive results—such as make, model, and color—within seconds by sending an image to the LLM. This approach is efficient for prototyping but poses potential challenges for scaling, as each scan incurs a cost associated with API usage and is vulnerable to fluctuations in service reliability. The application integrates a user-friendly interface with technical instrumentation via WildEdge, which tracks performance metrics like latency and confidence scores. In tests, the average response time was around 2.6 seconds, although this could increase in real-world scenarios with poor connectivity. While the generalist model demonstrates broad coverage, it also generates a high confidence score that may not always reflect true accuracy. The next steps include utilizing user feedback to enhance the dataset and ultimately developing an on-device classifier that could eliminate per-scan costs and reliance on external APIs, leading to faster, more reliable performance.