Member-only story
This is Part 1 of a 4-part series: Build → Harden → Observe → Ship. This post covers the full system — every component, how they connect, and why. Parts 2–4 go deep on security/guardrails, MCP + LangSmith observability, and the AWS deployment, respectively.
Most fraud-detection write-ups end at a confusion matrix. A model gets trained, ROC-AUC gets reported, and the story stops right where the interesting part begins: what happens when that score has to become a decision, in real time, inside a system a human can trust, audit, and interrupt.
This series documents a system built to answer that question properly — not a notebook, a running service. Three cooperating agents score a transaction from independent signals, a LangGraph workflow orchestrates them with a genuine human-in-the-loop (not a spinner), every stage is wrapped in guardrails, every run is traced in LangSmith, and the whole thing is containerized and deployed on AWS behind a FastAPI backend with a Streamlit frontend.
Repo: github.com/nitsourish/Agentic-AI-Orchestration — -Real-Time-Transaction-Fraud-Detection
The two architecture diagrams referenced throughout this post live in the repo at architechture/architecture.html
(full system) and…