# Architecture Summary

## Overview

BitMind operates as a distributed AI content detection platform with multiple interconnected components working together to provide accurate, scalable, and privacy-compliant deepfake detection services.

## Architecture

### Architecture Flow

The BitMind platform operates through five interconnected layers:

1. **User Layer**: Multi-platform interfaces (web, mobile, browser extension, enterprise clients) provide seamless access
2. **API Gateway**: Routes requests through specialized endpoints for platform, subnet, and enterprise operations
3. **Infrastructure**: Leverages Modal for ML inference, Supabase for data management, and Cloudflare for global distribution
4. **Bittensor Network**: Powers decentralized AI with Subnet 34 for fraud detection, distributed miners, and validator nodes
5. **Storage Layer**: Manages models and datasets across R2, Modal Volumes, and HuggingFace repositories

## Core Components

### 1. AI Infrastructure (Bittensor Subnet 34)

**Purpose**: Decentralized AI (DeAI) network that realizes a continuous, dynamic, incentivized competition for producing accurate detection models and novel, high-signal training data.

**Architecture**:

* **Detection Miners**: Submit packaged models (no live endpoints) for validator-side evaluation on curated benchmarks
* **Generation Miners**: Adversarially generate high-signal synthetic media that evolves the benchmark
* **Validators**: Evaluate submitted models locally, track benchmark versions, and distribute rewards based on measured performance
* **Adversarial Loop**: Generators push realism; discriminators improve detection — both co-evolve under aligned incentives
* **Model Management**: Automated submission, testing, and deployment flows

**Key Features**:

* Fully-generalized detection (works across different generative AI models)
* Continuous evolution with new generative AI
* Economic incentives for high-quality models
* Centralized evaluation inside validators improves privacy, determinism, and reproducibility
* Open, renewable training data via shared public corpus

### 2. Infrastructure Services

| Platform                                           | ML Inference                        | API Cloud Services                           |
| -------------------------------------------------- | ----------------------------------- | -------------------------------------------- |
| Admin at [app.bitmind.ai](https://app.bitmind.ai/) | High-performance ONNX inference     | Public API layer exposing endpoints to users |
| Developer dashboard and miner performance stats    | Cost-efficient model serving        | Image and video detection                    |
| API key management and usage analytics             | Smart GPU cache with auto-recovery  | Video preprocessing and segmentation         |
| Account administration and billing                 | Processing for efficiency           | C2PA integration for content provenance      |
| Integration with Bittensor ecosystem               | Security validation for ONNX models | High-throughput processing                   |

### 3. API Services

**Purpose**: Detection interfaces for consumer and enterprise workloads.

**Architecture**:

* **Subnet API** (`api.bitmind.ai`): Image and video detection with analytics, caching, and storage
* **Enterprise API** (`enterprise.bitmind.ai`): Same detection pipeline with zero data retention
* **Security**: API key auth, per-client rate limiting, Redis-backed quota tracking

**Key Features**:

* Zero data retention (enterprise)
* C2PA content credential extraction
* Similarity matching against known AI content
* Enterprise compliance (GDPR-ready)

### 4. Consumer Applications

**Purpose**: End-user experiences for fast, trustworthy deepfake detection.

**Architecture**:

* **Mobile (iOS/Android)**: Native apps for camera/gallery input
* **Web (bitmind.ai)**: Uploads, URL-based analysis, direct image address
* **Browser Extension**: On-page scanning and one-click verification across major sites

**Key Features**:

* Instant detection UX (mobile/web)
* URL/direct-address analysis (web)
* Privacy-first design (extension)
* Push notifications and usage insights (mobile)

## Architectural Highlights

### Detection Capabilities

* **Multi-Modal**: Images and videos
* **Real-Time**: Sub-second inference for images
* **Accuracy**: State-of-the-art detection performance
* **Adversarial Benchmarking**: Continuous refresh from generator submissions to surface new edge cases

### Privacy & Compliance

* **Zero Data Retention**: No input data stored (enterprise)
* **GDPR Compliant**: Zero data retention aligns with data minimization
* **SOC2**: Via Modal infrastructure ([details](https://modal.com/blog/soc2type2))
* **Reproducible Evaluation**: Results pinned to benchmark versions and model hashes

### Scalability

* **Horizontal Scaling**: Auto-scaling via Modal infrastructure
* **Global CDN**: Low-latency worldwide access via Cloudflare
* **Load Balancing**: High availability across regions