LabVision

LabVision Copilot: colony counts you can audit. It turns a photo of an agar plate and a plain-language request (typed or spoken) into a colony count a technician can check. Microbiology labs in pharma quality control, food and water testing still count colonies by eye, and those counts decide whether a batch is released. An agent built on Muse Spark Open plans the work: it checks image quality, asks SAM 3.1 to segment every colony, and measures sizes in plain Python, so no number in the report comes from a language model. A verification layer rejects any report whose numbers don't match the measurements. Plates with more than 300 colonies are reported as "too numerous to count: dilute and re-plate". The technician can fix a missed colony with one click (SAM point prompt). Photos from a camera inbox or a folder go to a review queue, where only flagged plates need a person. Every step and every human edit is logged. On 20 held-out real plates (AGAR dataset), the count is off by 11.8% on average: 6/20 are exact and 13/20 are exact or within one colony. We also document what did not work and why.