The problem Neural Torrent solves

Neural Torrent addresses the basic challenge of how AI agents can exchange training data in a fair and autonomous manner in a decentralized system. Currently, high-quality AI training data is costly, centralized, and dominated by a few large players. There is no open infrastructure that enables independent agents or models to find useful datasets, trade them with each other, and get legitimate credit for their contributions. In decentralized systems, agents typically download valuable data without contributing anything in return, causing a free-rider problem that undermines the entire network.
Neural Torrent fills this gap by merging the incentive dynamics of closed BitTorrent trackers with cryptographic reputation on-chain. Each AI participant enters the system with a wallet identity, contributes datasets like video or audio files for AI model fine-tuning, and accumulates a contribution score depending on the ratio of uploaded versus downloaded data. Each unit of data transferred is accompanied by a signed cryptographic receipt, and the upload and download metrics are maintained on-chain. Contributing participants get more access, and those who simply download without seeding get limited access until they pay back by contributing. This establishes a self-reinforcing incentive system where participation is visible, trackable, and incentive-compatible.
By doing so, Neural Torrent creates infrastructure for a decentralized AI data exchange layer, enabling agents to improve themselves through cooperative data sharing rather than relying on centralized brokers. Reputation becomes persistent and portable across tracker instances, ensuring that contribution history survives migrations or downtime. In essence, Neural Torrent solves the trust, incentive, and verification challenges in decentralized AI data sharing, enabling autonomous model self-evolution backed by cryptographic proof.

Challenges we ran into

Our team brought together people with very different levels of experience, ranging from researchers and developers to students who were new to distributed systems. This created an initial knowledge gap, particularly around complex concepts such as BitTorrent-style peer-to-peer networks. To overcome this, we leaned heavily on teammate-to-teammate teaching. More experienced members helped others build an intuition for how BitTorrent works—its incentive mechanisms, file sharing model, and network dynamics—so that the entire team could contribute meaningfully.
Another major challenge was direction and scope. We initially focused on improving the BitTorrent network itself, grounding our work in research and attempting to translate those ideas into a real MVP. After building much of the backend infrastructure, we realized that while improving BitTorrent was interesting, it was not the core problem we wanted to solve. This led to a key insight. The true strength of BitTorrent is not the protocol itself, but its ability to efficiently distribute large files in a decentralized way. That realization allowed us to connect our work to a much larger and more relevant problem: AI training data. By reframing training data as a large-scale, shareable asset—similar to files in a BitTorrent network—we aligned our technical foundation with a real-world need. This shift transformed our project from improving a network into applying BitTorrent economics to decentralize and scale access to AI training data.

Use of AI tools and agents

Over the course of the hackathon, we used GitHub Copilot and Claude to assist in developing Neural Torrent Network. Copilot helped generate key backend components including API routes, the /agent-tools discovery endpoint, CORS configuration, and cryptographic receipt logic.