FastAPI backend and evaluation layer for an explainable AI credit decisioning system. Parallel async LangGraph workflow (Risk, Fairness, Trajectory agents). Hybrid RAG: dense + BM25 sparse with RRF fusion over synthetic credit cases across multiple Qdrant collections, with substantial MRR improvement over baseline. BCT Regulation Agent RAG over a full-length banking law PDF, with a high pass rate and strong MRR and Recall@10 versus the baseline. Golden eval dataset with LLM-as-judge scoring.