The Problem

Most AI tutorials stop at "call the LLM API." This independent study asked a harder question: what does the full architectural journey look like — from a single LLM doing text extraction to a production multi-agent system that reasons, self-verifies, and runs on a schedule?

What I Built

• Designed and implemented 7 progressive prototypes showing the full evolution: DSPy typed-signature LLM extraction → RAG with ChromaDB (714 chunks, 768-dim embeddings) → FastAPI service layer → n8n scheduled orchestration → OutputCritic self-verification agent → ReAct agent with force-finish safeguards → production control tower with ML sustainability scoring
• Expanded corpus iteratively from 6 to 35 documents across 15 U.S. states and 7 major cities (1.89M characters); ML scoring model improved 63% in LOO MAE (13.51 → 5.01) as training data grew
• Final system passed 12/12 production checkpoint checks — FastAPI, Streamlit dashboard, n8n workflows, RAG retrieval, agent reasoning, and ML scoring all running end-to-end in Docker

Impact

A complete, documented study of how to architect and incrementally evolve an agentic AI system — from prompt engineering fundamentals to production multi-agent orchestration with tool-use, self-verification, and scheduled automation.