My name is Abhishek Roy, based in Kishanganj, Bihar. When building local AI agents, one major problem stands out: How do you route low-complexity prompts to fast, cheap models while preserving deep multi-pass reasoning for complex engineering prompts?
In this article, I’m walking through the design of TO-MoE, a local terminal-based Python prototype built on LangGraph.
Key Technical Sections:
Why Keyword Routing Fails (RFC Trap A vs Trap B)
Semantic Coordinate Routing with Cosine Similarity
Implementing Action Verb Dominance (AVD) & 1-Token Jury
State Protection: Silent Escalation & Quarantine Gates
Memory Architecture: ChromaDB Vector Store + Context-Assist Semantic Cache
Benchmark Results & Lessons Learned
Top comments (0)