AIForge
An open platform for sharing, downloading, and running AI agents and machine learning models locally — a developer-first model marketplace.
Problem
The ML community lacks a unified, open platform where developers can publish models and users can discover and run them locally without vendor lock-in. Existing solutions either require cloud execution or lack a cohesive developer experience for model sharing and local deployment.
Solution
AIForge is designed as an open marketplace with a developer-first philosophy. Users can publish models with metadata, download them via CLI or SDK, and execute them locally inside isolated Docker containers:
- ›Django REST backend with JWT authentication and model registry
- ›React/Vite frontend for browsing, searching, and publishing models
- ›Containerized model execution for isolation and reproducibility
- ›Python SDK for programmatic model discovery and local execution
- ›S3-compatible storage for model artifacts and versioning
Architecture
# AIForge Platform Architecture
┌──────────────────────────────────────────────────┐
│ Frontend (React/Vite) │
│ │
│ Browse Models ──► Download ──► Run Locally │
│ Publish Model ──► Upload ──► Set Metadata │
└────────────────────────┬─────────────────────────┘
│ REST API
▼
┌──────────────────────────────────────────────────┐
│ Backend API (Django) │
│ │
│ Auth ──► Model Registry ──► Storage ──► Jobs │
└───────────┬──────────────────────┬───────────────┘
│ │
▼ ▼
┌───────────────────┐ ┌──────────────────────┐
│ Model Hosting │ │ File Storage │
│ Server │ │ (S3 / local) │
│ │ │ │
│ Container each │ │ .pkl, .pt, .onnx │
│ model version │ │ model weights │
└───────────────────┘ └──────────────────────┘
│
▼
┌───────────────────┐
│ Local Execution │
│ Environment │
│ │
│ CLI / Python SDK │
│ Docker runner │
└───────────────────┘
Tech Stack
Django
Backend API & ORM
React + Vite
Frontend marketplace
Docker
Model execution isolation
PostgreSQL
Registry & user data
Django REST
API framework
S3
Model artifact storage
Challenges
Local Execution Isolation
Running arbitrary model code safely required containerization with resource limits and network isolation per execution.
Model Format Diversity
Supporting .pt, .pkl, .onnx and other formats required a unified interface abstraction with format-specific runners.
Dependency Management
Each model has unique Python dependencies. Solved by storing requirements.txt per model and building images on-demand.