AI Data Technician: An Expert-Guided Program Synthesis Agent for Scientific Data Processing

Across scientific domains, turning raw measurements into analysis-ready data remains a manual and subjective bottleneck. Much of this scientific data processing is still done by hand, as scientists inspect, clean, and curate datasets before any analysis can begin. In neuroscience, this often means scanning sensor recordings to separate clean neural signal from noise and artifacts. Learned artifact classifiers can reduce this labor, but they often replace expert judgment with an opaque decision function. We introduce AI Data Technician (ADT), an LLM-based program-synthesis agent that turns scientists' labels and natural-language feedback into executable feature programs and symbolic cleaning rules. ADT uses expert feedback to guide its search: it proposes feature programs, renders plots to inspect the data visually, fits compact one-vs-rest detectors, and composes them into an ordered multiclass decision list. Across two neural modalities, scalp electroencephalography (EEG) and intracranial EEG (iEEG), ADT exceeds or roughly matches strong baselines on signal classification tasks used for neural data cleaning, including non-interpretable methods such as random forests and deep neural networks. These results suggest that expert-guided LLM program synthesis can improve low-label neural signal cleaning, combining scientists' language feedback with the agent's own visual inspection of the data to produce executable rules that scientists can inspect, audit, and revise.