> For the complete documentation index, see [llms.txt](https://dxai.gitbook.io/dxai-light-paper/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://dxai.gitbook.io/dxai-light-paper/vision/ai-stack.md).

# AI stack

DXAI leverages a specialized **AI stack** that is purpose-built for medical diagnostics, combining advanced architectures, domain-specific training, and scalable deployment. Unlike general-purpose large language models (LLMs), our platform focuses on accuracy, interpretability, and efficiency in healthcare-specific applications.

**Key Components of the DXAI AI Stack:**

1. **Model Architectures:**
   * **Vision Transformers (ViTs)** and **CNNs** for medical image analysis.
   * Fine-tuned **LLMs** (e.g., Llava-Med, Mistral) for text-based diagnostics and reasoning.
2. **Domain-Specific Training:**
   * Trained on curated datasets such as SinoCT (CT scans), ISIC (dermatological images), and MURA (radiographs).
   * Specialized tuning ensures higher performance in sensitivity, specificity, and precision compared to generalized models.
3. **Real-Time Capabilities:**
   * Scalable architecture for processing thousands of cases daily.
   * Decentralized integration using $DXAI tokens for seamless access.

**Advantages of DXAI Over Commercial LLMs:**

* **Precision**: Domain-specific models outperform general LLMs in medical diagnostics by up to 20%.
* **Efficiency**: Optimized for medical tasks, reducing computational overhead.
* **Interpretable Outputs**: Provides heatmaps, overlays, and context-aware reasoning for clinicians.
* **Decentralization**: Blockchain-backed transparency and incentivization.

In a nutshell, considering commercial LLMs, this is the difference between our implementations and commercial LLMs:

<figure><img src="https://3521714503-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FksMt7szTV2lt6WZ21zX3%2Fuploads%2F8Q70w9qJMZWr9UwdTRnd%2FUntitled%20design.png?alt=media&amp;token=a02b2d88-ed65-4238-8636-570186510144" alt=""><figcaption></figcaption></figure>