The problem
Imaging equipment downtime is one of the biggest barriers to care in low- and middle-income countries, often because specialised biomedical engineering support is hard to reach.
What the team built
Guided by a survey across nine low- and middle-income countries, the team curated technical manuals for MRI and ultrasound systems into INGENZI_DatasetV1: 10,294 filtered question-answer-context pairs. They then fine-tuned Google's MedGemma-4b-it model with QLoRA, a parameter-efficient method, to read system error logs and produce step-by-step repair instructions.
Results
Compared with the base model, the fine-tuned model scored better on every metric:
- F1 score rose from 0.22 to 0.38
- ROUGE-2 rose from 0.18 to 0.41
- BERTScore F1 rose from 0.86 to 0.91
In practice, its answers to new troubleshooting questions are more precise and closer to the correct procedure. That is the foundation for the AI assistance in Ingenzi.