"My local model called a flagged result 'within range', so I stopped letting it do arithmetic"

When I first asked a local model to explain a lab report, it provided a result that was flagged as high within the reference range. Upon retesting, the model gave a different answer. The tool was intended to simplify understanding medical documents, but it was occasionally incorrect about the most critical aspect. The goal of this project is to ensure the model does not make decisions where accuracy is paramount.

The application takes a document's photo, PDF, or text input and provides a plain-language explanation. For lab reports, it summarizes results that fall outside the reference range and highlights key points. The tool also includes a glossary of terms and asks pertinent questions for the doctor. All this information is available in Hindi as well.

The application runs entirely on the user's laptop using a local model, without sending any data to external servers. The hosted version, however, sends data to a server and an AI provider, which should only be used with fictional sample reports. The code for this project is available on GitHub and is licensed under the MIT license.

The model used for this application is Gemma 3 4B, which runs on a local CPU without GPU support. An alternative model, Qwen 2.5 3B, was also tested, but its Hindi translation was found to be nonsensical. The application underwent testing with Bhavya, a classmate, who provided feedback on using the local version on an Android phone over Wi-Fi.