SentinAI is an autonomous agent for Windows that intelligently manages your storage using a hybrid AI + RAG + heuristics approach. It combines fast rule-based analysis with Phi-4 Mini AI and Weaviate vector memory for making safe, context-aware cleanup decisions.

• Hybrid Intelligence: Heuristics provide fast context, RAG retrieves past decisions, AI makes the final call
• Memory-Augmented: Weaviate vector database stores and retrieves past cleanup decisions for consistency
• Safe by Design: "Propose-Verify-Execute" pattern prevents accidental data loss
• Local & Private: All AI inference and vector storage runs locally - your data never leaves your machine
• User-Friendly: Simplified Home page for novice users, advanced Dashboard for power users

The application consists of multiple components:

1. HEURISTIC ANALYSIS (fast, rule-based)
   ├── Path-based detection (temp folders, caches, node_modules)
   ├── File pattern matching (*.tmp, *.log, *.cache)
   └── Winapp2 rules matching

2. RAG MEMORY RETRIEVAL (Weaviate + Ollama)
   ├── Generate embedding for current folder context
   ├── Query similar past decisions from vector store
   └── Include relevant memories in AI prompt

3. AI DECISION (Phi-4 Mini)
   ├── Receives heuristic context + RAG memories
   ├── Analyzes folder + files with historical context
   └── Makes FINAL safe/unsafe decision

4. MEMORY STORAGE
   ├── Store decision in Weaviate for future reference
   └── Build institutional knowledge over time

5. OUTPUT
   └── JSON response with confidence score

┌─────────────────────────────────────────────────────────────────┐
│                        SentinAI Web UI                          │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐  │
│  │    Home     │  │  Dashboard  │  │   Settings/Scheduler    │  │
│  │ (One-Click) │  │ (Advanced)  │  │   (Auto Cleanup)        │  │
│  └─────────────┘  └─────────────┘  └─────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                      AgentBrain Service                         │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐  │
│  │ Heuristics  │  │  Phi-4     │  │    RAG Memory Store     │  │
│  │   Engine    │──│  ONNX AI    │──│  (Weaviate + Ollama)    │  │
│  └─────────────┘  └─────────────┘  └─────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼
┌─────────────────────────────────────────────────────────────────┐
│                    Sentinel Service (gRPC)                      │
│  ┌─────────────┐  ┌─────────────┐  ┌─────────────────────────┐  │
│  │    USN      │  │   Cleanup   │  │   State Machine         │  │
│  │  Journal    │  │  Executor   │  │   Orchestrator          │  │
│  └─────────────┘  └─────────────┘  └─────────────────────────┘  │
└─────────────────────────────────────────────────────────────────┘

• Phi-4 Mini Instruct - Microsoft's powerful compact LLM with enhanced reasoning (~2.5GB)
• CPU or DirectML (GPU) - Configurable execution provider
• Hybrid approach - Heuristics validate, RAG retrieves context, AI decides
• Structured output - JSON responses for reliable parsing

• Weaviate Vector Database - Local vector storage at localhost:8080
• Ollama Embeddings - nomic-embed-text model at localhost:11434
• Contextual Recall - Retrieves similar past decisions to inform new ones
• Learning Over Time - Builds institutional knowledge from user decisions

• Windows/User temp folders → Safe
• Browser caches → Safe
• node_modules → Safe (developer confirmation)
• Build artifacts (bin/obj) → Safe
• Downloads with documents → Review Required

• One-Click Quick Scan - Simple scanning for novice users
• AI Auto Mode - Automatically cleans high-confidence items using AI + RAG
• Individual Item Approval - Approve or skip each item separately
• Bulk Actions - Approve All / Dismiss All for multiple items
• Real-time Progress - Visual feedback during scanning
• Smart Status - Shows pending items on page load

• Detailed Analysis - Full breakdown of all suggestions
• AI Reasoning - See why each item was flagged
• Confidence Scores - View AI confidence levels for each decision
• Category Filtering - Filter by safe/review status
• Execution History - Track past cleanup operations

• Quick Toggle Cards - Daily, Weekly, Monthly presets
• Friendly Time Picker - No cron expressions needed
• Safe Items Only - Auto-cleanup only affects pre-approved categories
• AI Confidence Thresholds - Configure minimum confidence for auto-cleaning
• Notification Options - Get notified before/after cleanup

• Community-maintained cleanup rules
• Auto-download from official source
• Grounds AI decisions in proven patterns

SentinAI/
├── src/
│   ├── SentinAI.Web/                  # Blazor Server Dashboard
│   │   ├── Components/
│   │   │   └── Pages/
│   │   │       ├── Home.razor         # User-friendly scan page
│   │   │       ├── Dashboard.razor    # Advanced analysis view
│   │   │       ├── Scheduler.razor    # Auto cleanup scheduling
│   │   │       └── Settings.razor     # Configuration page
│   │   ├── Controllers/               # API controllers
│   │   ├── Services/
│   │   │   ├── AgentBrain.cs          # Hybrid AI + RAG engine
│   │   │   ├── RagMemoryStore.cs      # Weaviate integration
│   │   │   ├── BrainConfiguration.cs  # CPU/DirectML config
│   │   │   ├── ModelDownloadService.cs
│   │   │   └── BrainInitializationService.cs
│   │   └── appsettings.json           # Configuration
│   │
│   ├── SentinAI.SentinelService/      # Background service
│   │   ├── Services/
│   │   │   ├── DriveMonitor.cs        # USN Journal monitoring
│   │   │   ├── UsnJournalReader.cs    # P/Invoke USN reader
│   │   │   ├── CleanupExecutor.cs     # Safe file deletion
│   │   │   └── StateMachineOrchestrator.cs
│   │   └── Program.cs
│   │
│   ├── SentinAI.Shared/               # Shared models
│   │   ├── Models/
│   │   └── Protos/agent.proto         # gRPC definitions
│   │
│   └── SentinAI.Brain/                # WinUI 3 app (optional)
│
├── download-models.ps1                # Manual model download script
├── build.ps1                          # Build script
└── install-service.ps1                # Service installer

• Windows 10/11 (22H2 or later)
• .NET 8 SDK or later - Download
• ~3GB disk space for AI model
• Docker (optional, for Weaviate RAG)
• Ollama (optional, for embeddings)

git clone https://github.com/gamepop/SentinAI.git
cd SentinAI
dotnet build SentinAI.sln

The Phi-4 Mini model (~2.5GB) downloads automatically on first run, or manually:

# Download CPU model (recommended)
.\download-models.ps1 -Provider CPU

# Or download DirectML (GPU) model
.\download-models.ps1 -Provider DirectML

# Or download both
.\download-models.ps1 -Provider Both

# Start Weaviate vector database
docker run -d --name weaviate -p 8080:8080 -p 50051:50051 `
  -e AUTHENTICATION_ANONYMOUS_ACCESS_ENABLED=true `
  -e PERSISTENCE_DATA_PATH=/var/lib/weaviate `
  -e DEFAULT_VECTORIZER_MODULE=none `
  -e CLUSTER_HOSTNAME=node1 `
  cr.weaviate.io/semitechnologies/weaviate:1.28.4

# Install and start Ollama for embeddings
winget install Ollama.Ollama
ollama pull nomic-embed-text

cd src/SentinAI.Web
dotnet run

Open http://localhost:5203 in your browser.

Edit src/SentinAI.Web/appsettings.json:

{
  "Brain": {
    "ExecutionProvider": "CPU",
    "ModelPath": "",
    "ForceModelRedownload": false,
    "InferenceTimeoutSeconds": 60,
    "MaxSequenceLength": 4096,
    "MaxOutputTokens": 150,
    "Temperature": 0.1
  },
  "Rag": {
    "Enabled": true,
    "WeaviateUrl": "http://localhost:8080",
    "OllamaUrl": "http://localhost:11434",
    "EmbeddingModel": "nomic-embed-text",
    "MaxMemories": 5
  }
}

Models are stored in:

• CPU: %LocalAppData%\SentinAI\Models\Phi4-Mini-CPU\
• DirectML: %LocalAppData%\SentinAI\Models\Phi4-Mini-DirectML\

• Local Processing: All AI runs locally, no cloud API calls
• No Data Collection: Files are analyzed but never uploaded
• Safe Defaults: Unknown folders default to "not safe to delete"
• User Confirmation: Ambiguous items require manual approval
• Individual Control: Approve/reject each cleanup suggestion separately

dotnet test

Enable detailed AI logs in appsettings.json:

{
  "Logging": {
    "LogLevel": {
      "SentinAI.Web.Services.AgentBrain": "Debug",
      "SentinAI.Web.Services.RagMemoryStore": "Debug"
    }
  }
}

.\build.ps1 -Configuration Release

• Hybrid AI + Heuristics engine
• CPU and DirectML support
• Phi-4 Mini integration
• Winapp2 rules parser
• Web dashboard
• RAG memory system (Weaviate + Ollama)
• User-friendly Home page with one-click scan
• AI Auto Mode with confidence thresholds
• Individual item approval/rejection
• Auto cleanup scheduler
• Settings page
• Real-time USN Journal monitoring
• Duplicate file detection
• Windows Store submission

MIT License - See LICENSE for details.

Dependencies:

• ONNX Runtime GenAI - MIT
• Phi-4 Mini - MIT
• Winapp2 - CC BY-NC-SA
• Weaviate - BSD-3-Clause
• Ollama - MIT

See CONTRIBUTING.md for guidelines.

• Issues: GitHub Issues
• Discussions: GitHub Discussions

Disclaimer: This software is provided "as is", without warranty of any kind, express or implied, including but not limited to the warranties of merchantability, fitness for a particular purpose and noninfringement. In no event shall the authors or copyright holders be liable for any claim, damages or other liability, whether in an action of contract, tort or otherwise, arising from, out of or in connection with the software or the use or other dealings in the software. Use at your own risk.

Built with .NET 8, Blazor, ONNX Runtime, Phi-4 Mini AI, Weaviate, and Ollama