Search for a command to run...
Decentralized Finance (DeFi) introduces open, permissionless, and censorship-resistant financial infrastructure that enables users to interact directly with digital assets without relying on centralized institutions. Despite these advantages, mainstream adoption remains limited due to a significant usability gap. Users are required to understand and interact with low-level constructs such as hexadecimal wallet addresses, volatile gas fee mechanisms, multi-step token approval flows, and network-specific transaction semantics, all of which demand technical expertise. This paper presents Vetrix, a conversational Web3 assistant designed to bridge this interaction gap through natural language interfaces. Vetrix leverages Large Language Model (LLM) based natural language processing to convert free-form user input into structured transaction intents. These intents are generated using Google Gemini 2.0 Flash and validated through a runtime Zod schema enforcement layer, ensuring that all outputs conform to predefined structural and safety constraints. The system then orchestrates transaction construction and confirmation through a non-custodial Verify-Then-Sign architecture, maintaining full user control over execution. The system supports multi-turn conversational gap filling, enabling incremental parameter collection, along with real-time gas estimation, AI-assisted risk detection, and portfolio balance retrieval. By combining probabilistic intent parsing with deterministic validation and execution, Vetrix eliminates the need for users to understand blockchain-specific details while preserving security guarantees. The architecture enforces strict separation between AI inference and transaction execution. The backend operates in a read-only capacity and does not access or store private keys. All cryptographic signing is performed locally on the user's device via MetaMask. The system is implemented on the Ethereum Sepolia testnet. Empirical evaluation across 45 test cases demonstrates 100% accuracy on balance queries, an overall intent classification accuracy of 69%, and a system response rate of 100%, with a mean end-to-end latency of 658 ms. Vetrix demonstrates that intent-centric interaction models can significantly improve blockchain usability by allowing users to express goals in natural language while delegating execution complexity to a structured and secure system.