Challenges I Ran Into

Building DOBBY was technically ambitious and came with significant challenges. The first major hurdle was managing and orchestrating multiple AI agents on OnDemand. Creating six distinct agent personas (The Auditor, Yield Hunter, Strategist, Sniper, Curator, and Visionary) required careful prompt engineering to ensure each agent had the right context and tools registered. I had to design a system where each agent could access specific widgets while maintaining consistent behavior across different user queries. The OnDemand Actions API was powerful but required extensive testing to get the plugin registration working correctly so that the AI could dynamically invoke the right tools based on natural language input.
API integration was another constant challenge. Many of the free-tier APIs I wanted to use had inconsistent documentation or unexpected rate limits. For example, some endpoints would work in testing but fail in production due to CORS issues or authentication problems. The GoPlus security API occasionally returned malformed responses that broke my parsing logic, requiring robust error handling. DeFiLlama's API structure changed mid-development, forcing me to refactor several components. I spent considerable time building fallback mechanisms and caching strategies with Redis to handle API failures gracefully without breaking the user experience.
Deployment to Vercel presented its own complexities. Next.js 15 with the App Router had some quirks with server-side rendering and client-side state management that required careful component architecture. Managing environment variables across development and production while keeping API keys secure was tricky. The initial build failed multiple times due to dependency conflicts between different packages, particularly with the drag-and-drop library and the charting components. I had to optimize bundle sizes because Vercel has deployment size limits, which meant carefully analyzing which libraries were truly necessary and finding lighter alternatives where possible.
The modular dashboard architecture itself was conceptually challenging. Building a system where 62 widgets could be dynamically loaded, rearranged, and connected to AI agents required thinking through state management very carefully. I used Zustand to create isolated stores for each widget, but coordinating data flow between widgets, the AI layer, and the backend APIs required a lot of debugging. Ensuring that widget layouts persisted across sessions while keeping the code maintainable was harder than expected.
Performance optimization was critical since DOBBY aggregates data from 10+ external APIs simultaneously. I implemented parallel API calls, intelligent caching with Redis, and lazy loading for widgets to keep the dashboard responsive. However, balancing real-time data updates with API rate limits meant building a sophisticated queuing system that prioritizes critical data while batching less urgent requests. WebSocket connections for live price feeds added another layer of complexity in managing connection stability and handling reconnection logic when users switch networks or lose connectivity.
Cross-chain data consistency was surprisingly difficult. Different blockchains return data in different formats, with varying block times and finality guarantees. Normalizing Ethereum transaction data to match Solana's structure while maintaining accuracy required building adapter layers for each chain. Gas calculations across Layer 2s like Base, Arbitrum, and Optimism each have unique quirks that needed special handling.