compute_vault

Pod AI Assistant (Working Title)

Privacy-preserving AI assistant that connects to a Solid Pod (Personal Online Datastore), reads personal data with user consent, and returns useful insights while ensuring full user ownership and control.

---

Problem

Most AI assistants require users to upload private data (notes, documents, health logs, etc.) to centralized servers, leading to privacy risks, vendor lock-in, and potential misuse.

Solid Pods address data ownership, but there is currently no standard architecture for enabling powerful LLMs to use Pod data without compromising privacy.

---

Core Idea

• User data remains in their Solid Pod, not in application databases
• The AI assistant receives scoped, temporary permissions to access specific Pod resources (e.g., /notes/, /health/)
• A backend AI service (local, remote, or hybrid) processes this data
• Results (summaries, insights, extracted tasks) are written back to the user’s Pod

This reverses the traditional model: users own their data, and AI services operate on it with permission.

---

High-Level Architecture

Components

Solid Pod

• Stores user data (notes, files, logs, chat history)
• Uses Solid access control and WebID identity

Client Application (Web/Desktop)

• Authenticates via Solid-OIDC
• Grants fine-grained read/write permissions

AI Backend Service

• Fetches only permitted data using access tokens

• Runs LLM pipelines (summarization, Q&A, RAG)

• Writes outputs back into the Pod

  approaches three approaches for the project:

On-device AI – high privacy since data stays local, but limited by smaller models. Backend server AI – more powerful models and faster development, but raises trust and data exposure concerns. Hybrid approach – sensitive data processed locally, heavy tasks handled by server.

I feel the hybrid model is the most promising from a research perspective, balancing privacy and performance

---

flowchart TD
    A["User"] --> B["Client App - Web/Desktop"]
    B -->|OIDC Login| C["Solid Pod"]

    B -->|Request Access Token| C
    B -->|Scoped Permissions| C

    B --> D["AI Orchestrator Backend"]

    D -->|Fetch Allowed Data| C
    D --> E["Privacy Layer"]

    E -->|Edge Mode| F["Local LLM (Ollama / llama.cpp)"]
    E -->|Trusted Mode| G["Cloud LLM (Claude / OpenAI)"]
    E -->|Hybrid Mode| H["Local Preprocessing + Cloud LLM"]

    F --> D
    G --> D
    H --> D

    D -->|Write Results| C
    D --> B

Loading