Investigate Replacing Azure API Calls with On-Device VLM

[dkillough]

Multiple researchers at CHI suggested replacing the Azure API calls with an on-device model, which may greatly simplify setup without compromising output quality. However, we should investigate feasibility with current consumer devices as on-device models likely have significantly higher performance overhead, degrading VRSight's object detection performance.

[dkillough]

One potential candidate may be smolVLM (https://huggingface.co/blog/smolvlm) but let me know if anyone has better suggestions or have tried these in the past. Sorry I forgot who mentioned it to me during the CHI demo sessions!