Manual assembly of high-precision components remains a critical point in otherwise highly automated production systems, with quality and traceability depending strongly on operator performance and ad-hoc visual checks. This paper presents a vision-based supervisory concept for assembly operations implemented with TinyML on an STM32 microcontroller. The proposed approach uses a low-cost STM32H747I-DISCO board with an attached camera module to monitor selected assembly steps and evaluate their conformity with a pre-trained image classification model. A complete end-to-end workflow is described, starting from 4K image acquisition at the assembly workplace, dataset preparation, and model training in Teachable Machine, through export to TensorFlow Lite for Microcontrollers and conversion by STM32Cube.AI, up to deployment and execution on the target hardware. The embedded application performs real-time image capture, classification, confidence estimation and on-device visualization, while supporting optional local time-stamped logging of recognised operations (e.g., to SD card) for later analysis. A proof-of-concept study on a representative reducer/actuator assembly demonstrates that, under stable illumination and a constrained set of classes, the TinyML based supervisor can distinguish between key assembly states and provide interpretable feedback to engineers. The work is positioned as a practical building block towards lightweight, retrofittable digitalization of assembly workplaces and as a basis for further research on robust datasets, integration with plant IT systems, and deployment in safety- and security-constrained industrial environments.