Oscilla

warehouse machines like forklifts and conveyors can suddenly stop working. Many teams only repair machines after they break. Oscilla helps them find the right machine to repair at the right time. How does it work? the vibration data from each machine is checked by a python service and the service gives a score based on the machine's normal condition. An alert show when two bad readings happen in a row. The system also checks data quality and rejects flat or broken signals, so a sensor problem will not make a machine look healthy. A manager can ask the AI assistant why a machine is getting worse and what actions should be taken. The assistant uses Gemma 4 31B, running on AMD Token Factory. It can use tools to read live data. Before showing an answer, the system checks its citations and numbers. The assistant can only suggest actions, and a human must approve them. Critical cases also need human review. after that, the manager can create a work order. Only technicians can be assigned to these tasks, and each technician can only see their own tasks. The system has four roles: Technician, Manager, Admin, and Viewer. Database Row Level Security is used to control what each role can access. Testing and results We tested the scoring system using the public CWRU bearing dataset. It detected all 9 faults and gave no false alarms on 10 healthy samples. It also correctly identified the fault location in 5 out of 9 cases, with no wrong locations. However, this dataset is quite easy, so we consider these results an early test. The demo machine data is simulated, and the impact numbers are based on assumptions that users can change. Machine load and speed are not included in the model yet. Oscilla is built with Next.js, Supabase, a FastAPI scoring service, and Vercel. Its source code is public and uses the MIT license.