Back to Home

DecAI Network Architecture

DecAI operates as a decentralized Layer-2 blockchain network that processes AI computation requests. The architecture consists of four key actor types, each with specific roles, responsibilities, and economic incentives. This multi-actor system ensures redundancy, fault tolerance, and true decentralization while maintaining high performance and security standards.

All components interact through smart contracts deployed on the Layer-2 blockchain, which handles prompt assignment, consensus verification, and automated reward distribution. The network dynamically adjusts node selection based on performance metrics, geographic distribution, current load, and specialized capabilities.

⛓️

Layer-2 Blockchain (L2)

Central Infrastructure Layer

Core Functionality

The Layer-2 blockchain serves as the foundation of the DecAI network, providing security, scalability, and low transaction costs. It leverages the security of underlying networks like Ethereum or Avalanche while operating as an independent layer optimized for high-frequency AI prompt transactions.

Smart Contracts Handle:

  • Prompt Assignment: Randomly assigns incoming prompts to compute nodes using Verifiable Random Function (VRF)
  • Consensus Verification: Validates hash comparisons and consensus results from validator nodes
  • Reward Distribution: Automatically distributes DCI tokens to participating nodes based on their contributions
  • Reputation Management: Updates node reputation scores based on performance metrics
  • Slashing Mechanism: Enforces penalties for malicious behavior by reducing staked tokens
Transaction Cost
<$0.01
Confirmation Time
<5 seconds
Scalability
Millions/day

Compute Nodes

70% of Rewards

Primary Role

Compute nodes are the workhorses of the DecAI network. They execute AI prompts using their local GPU/CPU resources, processing requests in parallel to ensure fast response times. These nodes are selected randomly using VRF based on their stake, reputation, and specialization match.

Responsibilities:

  • Execute AI prompts locally using available computational resources
  • Generate hash results (SHA-256) of computation outputs
  • Provide digital signatures (ECDSA) for result verification
  • Report resource usage metadata (GPU time, memory, energy consumption)
  • Maintain high uptime and consensus match rates
Minimum Stake
10,000 DCI
Required Uptime
>95%
Consensus Match Rate
>90%
Specialization
LLM, Images, Embeddings

Reward Distribution

Compute nodes receive 70% of prompt execution fees, split proportionally based on execution speed (faster nodes get higher share), result quality (consensus match rate), and overall uptime/reliability. This incentivizes nodes to maintain high performance and accuracy.

Validator Nodes

20% of Rewards

Primary Role

Validator nodes verify the results submitted by compute nodes and participate in the consensus mechanism. They run lightweight verification algorithms to check result validity without re-executing the entire prompt, making the verification process efficient and scalable.

Responsibilities:

  • Verify hash results from compute nodes
  • Participate in consensus checks (≥2/3 match requirement)
  • Challenge suspicious results and trigger audit processes
  • Run lightweight verification algorithms
  • Maintain accuracy in challenge decisions
Minimum Stake
50,000 DCI
Security Level
High
Verification Type
Lightweight
Can Trigger
Audits

Reward Distribution

Validator nodes receive 20% of prompt fees, split based on the number of validations performed, accuracy of challenge decisions, and overall network contribution score. Higher accuracy in identifying malicious behavior results in higher rewards.

🔮

Oracle Nodes

Treasury Bonus

Primary Role

Oracle nodes provide subjective evaluation for non-deterministic AI outputs where direct hash comparison is impossible (e.g., text generation, image synthesis). They use advanced AI models to assess semantic similarity, quality, and coherence of results when compute nodes produce different but potentially valid outputs.

Evaluation Methods:

  • Semantic Similarity Metrics: Cosine similarity of embeddings (BERT, GPT), BLEU/ROUGE scores, CLIP similarity
  • Embedding Comparison: Multi-dimensional vector space analysis, clustering, outlier detection
  • Model-Based Scoring: Fine-tuned quality assessment models, RLHF-style human feedback integration
  • Multi-Oracle Consensus: 3+ oracles must agree for high-stakes validations
Minimum Stake
100,000 DCI
Reputation Required
>4.5/5.0
Accuracy Required
>95%
Governance
DAO or Trusted

Oracle Governance

Oracle nodes can be DAO-governed or run by trusted entities. Community votes on oracle decisions for high-stakes validations, with higher-reputation oracles having more influence. Oracles earn rewards from the treasury but face penalties for incorrect assessments, ensuring honest evaluation.

Audit Nodes

Slashing Rewards

Primary Role

Audit nodes act as final arbiters when consensus fails or disputes arise. They re-execute prompts independently to verify results and resolve conflicts. Selected from top-performing compute nodes with the highest reputation scores, they serve as the network's ultimate truth verification mechanism.

Responsibilities:

  • Re-execute prompts when consensus fails (<2/3 match)
  • Resolve disputes between compute nodes
  • Act as final arbiters in edge cases
  • Provide independent verification of suspicious results
  • Determine slashing penalties for malicious nodes
Selection Criteria
Top Reputation
Source
Best Compute Nodes
Reward Source
Slashing Penalties
Role
Final Arbiter

Economic Model

Audit nodes receive bonus rewards from slashing penalties collected from malicious nodes. This creates a self-sustaining security mechanism where dishonest behavior funds the verification system. The high reputation requirement ensures only the most reliable nodes can serve as auditors.

👥

Users

Network Consumers

User Interaction

Users interact with the DecAI network by submitting AI prompts through a simple interface. Each prompt is converted into a blockchain transaction containing the prompt data, model parameters, priority level, and payment in DCI tokens. Users receive verified results on-chain with full transparency about which nodes processed their request.

User Capabilities:

  • Submit AI prompts with specified models and parameters
  • Set priority levels (standard, high, urgent) affecting latency and cost
  • Pay with DCI tokens (dynamic pricing based on model complexity and network load)
  • Receive verified results with on-chain proof of execution
  • Query blockchain to verify which nodes processed their prompts
  • Access historical performance data and reputation scores
Base Cost
0.1 DCI
Standard Latency
10-30s
High Priority
5-15s
Transparency
100% On-Chain

Network Interaction Flow

Complete Request Lifecycle

  1. User Submission: User submits prompt with DCI payment to smart contract
  2. VRF Selection: Smart contract randomly selects 3-5 compute nodes using Verifiable Random Function
  3. Parallel Execution: Selected compute nodes execute the prompt simultaneously
  4. Hash Submission: Each compute node submits hash result, signature, and metadata
  5. Validator Verification: Validator nodes check hash comparisons and verify consensus
  6. Consensus Check: If ≥2/3 hashes match, consensus achieved. If not, Oracle/Audit evaluation
  7. On-Chain Storage: Verified hash and proofs stored permanently on blockchain
  8. Result Delivery: User receives output via encrypted channel
  9. Reward Distribution: Smart contract automatically distributes DCI tokens (70% compute, 20% validator, 10% treasury)
  10. Reputation Update: All participating nodes receive reputation score adjustments

Key Architectural Features

Decentralization
Thousands of independent nodes, no single point of failure
Transparency
Every execution recorded on-chain, fully auditable
Security
Multi-layer verification, slashing, reputation system
Scalability
Layer-2 architecture supports millions of daily prompts
Efficiency
10x cheaper than centralized alternatives
Flexibility
Supports any AI model, not locked to specific frameworks