🤝 Top Collaborators
.png&w=640&q=75)
.png&w=640&q=75)
🤓 Latest Submissions
Multilönd AI
A missing charger is easy to add while a box is open. Once the lid closes, the same mistake becomes rework or a customer return. In hand assembly and packing, the next step often hides the last one. Multilönd AI checks work at those visible moments. A station camera watches the bench; when the scene is steady, one frame goes to an open vision model with a checklist question that describes what correct looks like. Fixed rules, not the model, decide the result: Pass moves on, Fix it marks the problem with a red box, Retake asks for a clearer view. A step that covers earlier work (closing a lid or panel) stays locked until its checks pass or a person writes an override reason. Every attempt is saved with its photo, unit ID and time. The web app runs live on a camera or on a recorded video, and anyone can write their own checklist in plain lines, so a new product is a new checklist, not a new machine. Our demo is an accessory-box pack-out; the same engine also checks a PC build. AMD: inspections run on DeepSeek V4.1 Flash, an open-weight model, on AMD Token Factory. Because the model is open, it can run on a factory's own AMD GPU so product photos never leave the site. Results: on 34 MVTec AD cable photos (earlier run, Gemma 4 on AMD Token Factory), a generic question caught 2/24 defects; a question describing the correct cable caught 14/24, with zero false alarms on 10 clean photos and about 5.7 s per check. Ten defects still escaped, so broader validation is our next step.
18 Oct 2026
