CLARA: Democratizing plastic sorting through LLM- based agentic coding

Most European plastic sorting facilities, predominantly Small and Medium-sized Enterprises (SMEs), still rely on manual processes. Existing automated alternatives, typically specialized vision systems, are either closed solutions or require full reprogramming whenever the sorting flow changes, leaving most operators outside the reach of automation.

Classification & Advanced Recognition Automation (CLARA) is an edge-deployed sorting system that addresses both barriers: it is configured through a Large Language Model (LLM) rather than code, allowing any operator to adapt the sorting flow without expert intervention. At run time, a Mask Region-based Convolutional Neural Network (R-CNN) segments items as rigid or flexible and a Kalman filter tracks them; color and other visual features are matched against a rule set that routes the object to its target container. That rule set is generated by Agentic AI approach (three task-adapted LLMs) including an orchestrator that manages the operator's dialogue, an analyzer that extracts discriminative features from sample images, and a parser that configures the containers from textual descriptions.

Under real operating conditions, the system achieved a sustained throughput of 940 classified objects per hour (obj/h), exceeding the 800 obj/h industry target, with a mean detection confidence of 0.968 and Mask R-CNN achieved Mean Average Precision (mAP)@0.5 of 0.861. A non-technical operator completed full sorting-rule configuration in under three minutes through natural-language interaction, a task that previously required code-level reprogramming by a software engineer.

The outcomes of CLARA project show that the complexity and expertise barriers historically excluding SMEs from automated plastic sorting can be removed without compromising performance. Pairing a lightweight perception pipeline with a small-LLM based configuration generalizes beyond plastic sorting, offering a transferable framework for industrial automation in domains where operator expertise, not throughput, is the binding constraint.