Decentralized AI training platform — train models on your data, reward the world for helping.
TrainNet.ai is a decentralized platform for training machine learning models using volunteer compute resources.
It lets users:
- Upload their datasets
- Select a pre-configured model architecture
- Automatically distribute training to worker nodes
- Merge results into a final model
- Download trained weights or integrate via API
And it rewards contributors (workers) who train shards of the model.
- AI training is expensive, centralized, and cloud-locked.
- Millions of GPUs and CPUs around the world sit idle.
- TrainNet.ai connects compute with need — like mining, but actually useful.
- ✅ Task queue with Redis
- ✅ Docker-based worker agents (PyTorch)
- ✅ Dataset sharding & aggregation
- ✅ Finalizer with weight merging
- ✅ FastAPI backend + REST API
- ✅ React frontend (upload, status, download)
- ✅ Supabase for storage and metadata
- User uploads dataset via
/trainendpoint or React UI - Backend shards the data and creates Redis queues
- Distributed workers fetch shards and train models
- Finalizer checks for completion and merges models
- User downloads their trained model
- Add model verification / validation
- Reputation and reward system for workers
- Custom hyperparameter selection
- Token-based marketplace
- Federated / ZK training support (R&D)
trainnet-ai/
├── backend/ # FastAPI backend: API, task logic, file mgmt
├── worker/ # Python shard trainer (Dockerized)
├── finalizer/ # Weight aggregator
├── frontend/ # React client UI
├── deployments/ # Docker Compose configs
└── README.md # You're here
Ayaal Santaev — backend engineer, open-source enthusiast, and believer in a fairer, more decentralized AI infrastructure.
This is a solo-built MVP.
Currently seeking grant support, early adopters, and contributors.
📫 Contact: ayaalsantaev@gmail.com
📎 Pitch deck: TrainNet_PitchDeck.pdf
📁 GitHub: https://github.com/Galiusbro/TraiNet
Coming soon: Docker quickstart + demo mode.