QA Vibes: Rapid Deployment of an MPC Auditor Via LLM-Augmented Vibe Coding
Abstract
Purpose
Transitioning Linac QA from time based testing to data-driven quality management is hindered by the high technical barrier of developing custom clinical tools. This work demonstrates "vibe coding” as a viable solution, leveraging a Large Language Model (LLM) to rapidly architect, debug, and deploy a production-ready auditing platform for Varian's Machine Performance Check (MPC). The resulting platform utilizes Statistical Process Control (SPC) to identify risks obscured by traditional binary pass/fail reviews.
Methods
A clinical physicist with limited software engineering experience utilized an LLM (Gemini 3.0) to co-develop a Python-based auditing ecosystem. The LLM was used to architect a recursive scraper to parse longitudinal MPC data from network drives.The LLM further assisted in implementing a 5-sigma outlier filter, calculating Process Capability (Cp) and 30-day drift velocity for over 120 parameters. Total development time, from conceptualization to an interactive HTML "Traffic Light" dashboard, was quantified to assess the "speed-to-clinic" advantage of LLM-assisted coding.The LLM-generated logic was manually validated against a benchmark dataset.
Results
Total development time from conception to a working prototype was 3.7 hours. Automated analysis of the data revealed that ~85% of daily QA metrics (e.g., couch and jaw positioning) were "Six Sigma" stable (Cp > 2.0), suggesting that manual daily review for these parameters may be redundant. Conversely, the IsoCenterMVOffset exhibited a Cp of 0.70 (Incapable) with a positive drift of +0.0041 mm/day. This statistical failure was detected despite the system passing Varian tolerances, providing a proactive trigger for intervention before clinical thresholds were breached.
Conclusion
LLM-augmented custom coding development represents a paradigm shift in medical physics with the potential to move the field closer to real time predictive machine maintenance. This study demonstrates that "vibe coding" bridges the gap between complex clinical data and actionable quality management, allowing physicists to rapidly deploy tools that optimize QA workflows.