Building Lumina came with both technical and architectural challenges.

One major hurdle was *bridging prediction with real execution. While scheduling transactions in MongoDB was straightforward, nothing happened when the scheduled time arrived. The system lacked an active execution layer. This was solved by designing a *background worker (Executioner) that continuously monitors scheduled transactions and triggers execution automatically, creating a closed-loop system.

Another challenge was *handling volatile gas prices. Gas values change rapidly, making static logic unreliable. I overcame this by introducing a *machine learning–based time-series prediction approach, allowing the system to make data-driven decisions instead of fixed assumptions.

Finally, converting natural language user intent into structured blockchain logic was non-trivial. Integrating Gemini as a reasoning layer helped extract intent, constraints, and conditions accurately, enabling seamless automation without user confusion.

Each challenge pushed the project closer to a real-world, production-grade system rather than a simple prototype.