DeFlow

Problem Statement

DeFlow is an AI-driven DeFi automation platform that transforms natural language into executable blockchain workflows. Users describe what they want to do in plain English (e.g., "swap 0.01 ETH to USDC on Base"), and our AI agent system automatically plans, validates, and executes the multi-step transactions across 15+ EVM chains.The platform features:Natural language workflow builder powered by ASI:One agentsMulti-chain support (Ethereum, Base, Arbitrum, Optimism, Polygon, etc.)Automated swap execution via Uniswap V3WETH wrapping/unwrapping functionalityReal-time transaction notifications via BlockscoutPersistent workflow templates with MongoDBAccount abstraction using Lit Protocol for secure key managementVincent SDK authentication for seamless user onboardingMeTTa RAG knowledge base for intelligent DeFi decision-makingUsers can save workflows as templates, chain multiple operations together, and let the AI handle all the complexity of cross-chain DeFi interactions.

Solution

DeFlow uses a hybrid architecture combining traditional web2 components with cutting-edge AI and blockchain technology:Frontend (React + Vite):Vincent SDK for passwordless OAuth authenticationBlockscout SDK for real-time transaction notifications and explorer integrationVisual workflow builder with node-based UI for planning multi-step operationsBackend Services (Node.js + Python on GCP):Node.js Express API handling Vincent auth, workflow CRUD operations, and blockchain interactionsPython Flask service running ASI:One agents with MeTTa reasoning enginePM2 process management with Google Cloud Ops Agent for centralized loggingMongoDB Atlas for workflow persistence and user dataBlockchain Integration:Lit Protocol PKPs (Programmable Key Pairs) for distributed key management and transaction signingEthers.js for multi-chain RPC connections across 15+ networks via AlchemyUniswap V3 SDK for optimal swap routing and price calculationsCustom token transfer and WETH wrapping utilitiesAI/Agent Layer:Fetch.ai's uAgents framework for autonomous agent orchestrationASI:One client for natural language understanding and workflow planningMeTTa (Meta Type Talk) reasoning engine for DeFi knowledge graphsAgentverse integration for agent deployment and mailbox communicationBlockscout MCP agent for blockchain data queryingHacky/Notable Solutions:Built custom MCP (Model Context Protocol) client to bridge ASI agents with blockchain toolsImplemented BigInt serialization workarounds for handling wei amounts in JSON APIsCreated dynamic amount inheritance system for chained swaps (output of swap1 → input of swap2)Used Vincent's delegatee pattern to avoid storing user private keysIntegrated GitHub Actions for zero-downtime deployments to GCP's free tier (e2-micro VM)Engineered PM2 log collection with Ops Agent to get structured logs in GCP without containersThe most challenging part was getting ASI:One agents to reliably call blockchain functions with correct parameter types while maintaining natural language flexibility. We solved this with a hybrid approach: agents generate high-level plans, which get validated and typed by our Node.js backend before execution.