aceso-medgemma-1.5-4b-it-Q4_K_M

Project Aceso: MedGemma-1.5-4B-it (4-bit Q4_K_M)

This repository contains the optimized Tier 1 Intelligence Layer for Project Aceso, a tri-tier clinical assistant system developed for the Med-Gemma Impact Challenge.

🩺 Project Overview

Project Aceso is designed to streamline clinical documentation by converting real-time patient encounters into professional SBAR (Situation, Background, Assessment, Recommendation) reports. By utilizing on-device Edge-AI, Aceso ensures total data sovereignty and offline reliability in high-pressure medical environments.

🧠 Model Specifications

This model is a 4-bit quantization of Google's MedGemma-1.5-4B-it, specifically tuned for the hardware constraints of modern mobile devices like the Samsung S23.

• Quantization: Q4_K_M (Medium)
• Format: GGUF
• Size: ~2.3 GB (Optimized for 8GB RAM devices)
• Hardware Acceleration: Fully compatible with Vulkan-accelerated backends.

🛠️ Quantization Credits & Attribution

Special recognition is given to the Unsloth team for the quantization of this model.

• Tooling: This build was processed using Unsloth, utilizing their high-efficiency kernels to maintain clinical accuracy while significantly reducing the VRAM footprint.
• Methodology: The Q4_K_M method was selected to provide the optimal balance between speed and medical reasoning depth for the Project Aceso Tier 1 client.

🚀 Deployment in Project Aceso

This model serves as the local inference engine within the Aceso Android App.

• Privacy First: 100% of the inference happens on-device; no patient data leaves the phone.
• Hardware Aware: The app utilizes the Snapdragon 8 Gen 2's Adreno GPU via Vulkan for high-speed, on-device SBAR generation.
• Human-in-the-Loop: All AI outputs are treated as drafts and require authorization by a licensed healthcare professional before being logged to the Aceso Hub.

⚖️ Licensing

This model is released under the Health AI Developer Foundations license. Users must comply with all ethical and regulatory guidelines for medical AI as established by the competition organizers.