https://clip-on.ai/updates.html Field notes and progress from Clip-on.AI — AI-driven roadside litter analysis. Dataset milestones, model benchmarks, hardware builds. en-us Mon, 20 Jul 2026 00:00:00 -0800 https://clip-on.ai/updates.html#update-2026-03 https://clip-on.ai/updates.html#update-2026-03 Sun, 01 Mar 2026 00:00:00 -0800 An open-source two-tier system: on-device YOLO26 detection (0.905 mAP50) plus vision-language classification. 4,499 images, 18,464 labels, 71 litter categories across 14 cleanup sessions. Head-to-head benchmark of Gemini, GPT-4.1, and Claude; QLoRA fine-tune of Qwen3-VL-8B reaching 0.024 eval loss on a consumer GPU.