Spaces:
Sleeping
Sleeping
| title: MANN Engram Showcase | |
| emoji: 🧠 | |
| colorFrom: blue | |
| colorTo: indigo | |
| sdk: gradio | |
| sdk_version: 6.11.0 | |
| app_file: app.py | |
| pinned: true | |
| license: mit | |
| github: https://github.com/Mr-wuff/MANN-Engram | |
| tags: | |
| - Medical-AI | |
| - Multimodal | |
| - SiGLIP | |
| - Edge-Cloud | |
| - Routing | |
| - Privacy-Preserving | |
| # MANN-Engram: Edge-Cloud Multimodal Semantic Router | |
| 🧠 A privacy-first, zero-hallucination shield for clinical vision-language models. | |
| ## Overview | |
| MANN-Engram is an orchestration layer designed to solve the "Clinical Input Noise" problem in Large Multimodal Models (LMMs). It combines cloud intelligence for logical intent distillation with edge-side tensor routing for precise multimodal evidence selection. | |
| ## Core Philosophy | |
| Clinical patient data is often noisy. A single diagnostic session may contain emotional complaints, billing frustrations, and unrelated imaging scans. Downstream VLMs can suffer from semantic drift or hallucinations when they process this irrelevant context. | |
| MANN-Engram acts as a surgical filter: | |
| - **Linguistic Distillation (Cloud)**: Uses Qwen-2.5-72B to remove non-medical noise and extract purified clinical intent. | |
| - **Tensor Routing (Edge)**: Projects intent and imaging into a shared latent space via a skew-Gaussian optimized SiGLIP engine. | |
| - **Precision Pruning**: Only verified core evidence is passed to the VLM context window. | |
| ## Key Capabilities | |
| - **Linguistic de-noising**: Removes billing complaints, food issues, and emotional noise from patient narratives. | |
| - **Multimodal saliency routing**: Selects relevant diagnostic scans (MRI/CT/X-ray) from unordered data dumps. | |
| - **Dynamic gate control (`Top_p`)**: Allows clinician-controlled precision vs. recall. | |
| - **Edge-cloud synergy**: Keeps privacy-sensitive tensor routing local while offloading reasoning to the cloud. | |
| ## Benchmark Case: "The Neurological Decoy" | |
| - Scenario: A patient complains about hospital food and leg cramps, but also mentions a seizure and left-sided numbness. | |
| - Input pool: 1x Brain MRI (Tumor), 1x Chest CT, 1x Abdominal CT, 1x Leg Angiogram. | |
| - Challenge: Ignore the decoy complaints and identify the tumor. | |
| - Result: At `Top_p = 0.6`, MANN-Engram achieves 100% noise suppression and routes only the Brain MRI as core evidence. | |
| ## Quick Start Guide | |
| 1. Obtain a Hugging Face token. | |
| 2. Enter it in the Settings panel. | |
| 3. Paste a clinical narrative and upload multiple images. | |
| 4. Adjust `Top_p`: | |
| - `0.6` for high precision | |
| - `0.85` for clinical safety | |
| 5. Review the output dashboard for purified intent and routed evidence. | |
| ## Developer Integration | |
| The core logic is available as a standalone Python SDK. | |
| ```bash | |
| git clone https://github.com/Mr-wuff/MANN-Engram.git | |
| cd MANN-Engram | |
| pip install -r requirements.txt | |
| ``` | |
| ## Repository Structure | |
| - `mann_engram_en/` — Core routing and intent extraction logic. | |
| - `weights/` — Pre-trained skew-Gaussian weights for SiGLIP routing. | |
| - `examples/` — Jupyter notebooks for threshold sensitivity analysis. | |
| ## Citation | |
| If you use this project in research or clinical applications, please cite: | |
| ```bibtex | |
| @software{MANN_Engram_2026, | |
| author = {WuFeiFan}, | |
| title = {MANN-Engram: Edge-Cloud Multimodal Semantic Router}, | |
| url = {https://github.com/Mr-wuff/MANN-Engram}, | |
| year = {2026} | |
| } | |
| ``` | |
| Created with ❤️ by Mr-wuff. Focused on advancing trustworthy AI in healthcare. | |
| Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference | |