vbookshelf/Offline-MedAi-Console-TDA

Offline MedAi Console v4.0 TDA

A multimodal AI Collaborator for clinicians. Powered by MedGemma 4b.

Demo Video - v1.0
https://www.youtube.com/watch?v=X3-6MpZp89s

With the right AI partner, ordinary people can do extraordinary work.

AI is not perfect. It hallucinates. It gets things wrong. This lack of reliability limits its use as an autonomous medical tool.

However, when used as a collaborative tool, AI can add significant value without having to be perfect. AI brings speed and a vast knowledge base. The clinician brings medical judgment and an understanding of the patient’s entire context.

The MedAi Console is a multimodal AI collaborator for clinicians. It supports text, images, PDF files and DICOM files. And it runs completely offline.

The console is powered by Google's MedGemma 4B model in BF16 GGUF format. MedGemma is a multi-skilled medical model.
It can interpret medical images in four specialities:

• Radiology (X-rays, CT scans)
• Digital Pathology (Tissue slides)
• Dermatology (Skin conditions)
• Ophthalmology (Retinal scans)

The model can also understand medical documents. Moreover, it has the general purpose capabilities of the Gemma3 model.

This console solves two problems:

• Universal Access: It works where there's no internet access — disaster zones, rural clinics and developing countries.
• Regulatory Compliance: It sidesteps a U.S. regulatory headache. When doctors use cloud-based AI with patient data, HIPAA (Health Insurance Portability and Accountability Act) requires them to sign BAAs (Business Associate Agreements) with cloud providers. This introduces a complex ongoing compliance process. Also, it's expensive for small practices and medical non-profits. An offline tool eliminates the BAA requirement. There are still compliance requirements, but the administrative and legal burden is reduced.

What is a Thumb-Drive App (TDA)?

• A portable folder-based AI app that doesn’t need system wide installation.
• Runs completely offline.
• All dependencies are bundled - including Ollama and the model.
• Created to be run on a Silicon Mac.
• Easy to share using AirDrop or a thumb-drive.
• All dependencies travel with the copy.
• Can be run on an external SSD.
• Double-click to launch.

Designed to bring the power of AI to places where there's no internet.

How to Use

System Requirements:

• Operating System: MacOS
• Computer: Apple Silicon Mac (M1, M2, M3, M4)
• RAM: 16GB
• Free disk Space: 10.2 GB

1. Download the MedAi-Console-v4.0-TDA.zip file and unzip
2. Place the unzipped folder on your desktop. Then open the terminal and cd into the MedAi-Console-v4.0-TDA folder:

   cd Desktop/MedAi-Console-v4.0-TDA
   

3. MacOS often quarantines downloaded files. To make the launch script executable, paste this command into the terminal and press Enter:

cat start-mac-app.command > temp && mv temp start-mac-app.command && chmod +x start-mac-app.command

4. Open the MedAi-Console-v3.0-TDA folder and double-click this file: start-mac-app.command
   If a MacOS security popup appears, click: "Allow"

5. The chat window will open. This may take a few seconds. Please be patient if you don't see anything happening. During normal use the app will open much faster. Because the model needs to load, it may take a few seconds to get a response to your first message. Subsequent responses will be much faster.

One Command. Complete Control.

You would have noticed that, to make this app executable, this command needed to be typed in the terminal:

cat start-mac-app.command > temp && mv temp start-mac-app.command && chmod +x start-mac-app.command

To remove this requirement the start-mac-app.command file would need to be code signed. Most macOS apps distributed outside the App Store are code signed using an Apple Developer ID certificate. Code signing requires an active Apple Developer Program membership.

This project does not use code signing because it introduces a point of failure. If the Apple membership lapses, or if Apple revokes the certificate for any reason, signed apps will stop launching cleanly for new users. This risk is unacceptable.

In exchange for the minor inconvenience of typing one command into the terminal, you get a reliable app that has no dependency on any third party. You have complete control.

Known Issues

• Re pywebview: When using voice input and output, the user experience is not stable and robust. The mic keeps going on and off during voce input also, the volume and completeness of the voice ouput varies. This is not as good as simply opening the app in Chrome (or the user's default browser). The voice input/ouput when using the Chrome browser is much better. I did not address this error because it serves as a good example of how the user experience can vary depending on the architecture being used.

References

• Thumb Drive App Concept
  https://github.com/vbookshelf/Thumb-Drive-App-Concept

• Kaggle Writeup - v1.0
  https://www.kaggle.com/competitions/med-gemma-impact-challenge/writeups/offline-medai-console

• Offline MedAi Console v1.0
  https://github.com/vbookshelf/Offline-MedAi-Console

Revision History

Version 4.0
20-March-2026
Prototype. Released for testing, education and inspiration.

• App is now fully self-contained.
• Bundled the Python 3.12 interpreter.
• Used uv cache to bundle the wheels for all packages.
• Added a venv token system that deletes a venv that may have been shipped with the app.
• Added a setup mode in start-mac-app.command that auto bundles the packages.
• Replaced openai-whisper with whisper-mlx because openai-whisper could not be bundled - constant errors and many frustrations.
• Remove pywebview because it could not be bundled - proxy-tools error. App now runs in the browser. The terminal needs to be shut down in order to shut down the app. This is a downgrade compared to the version 3.0 user experience.
• Voice chat can be unstable. Removing STT and TTS would produce a simpler, lighter, more robust app. The user experience would also be better beacuse pywebview can be used to create a native macOS app experience.
• Overall this serves as a good example of how to use uv to bundle an app. However, uv does not have a download capabilty like "pip download." Having this capability would make packaging apps much less prone to errors.
• The main learning point after this experience: Put MedGemma into the "Ai-On-A-Stick" app architecture - simple, robust, secure, easy to package, better user experience.

Version 3.0
10-March-2026
Prototype. Released for testing, education and inspiration.

• Updated data encryption code
• Updated UI to use pywebview
• Updated auto launch process