Swiggy Nexus Case Study | Aryan Dani
A LangGraph agent that stages food, grocery, and dining actions and blocks every write until a human approves.
Problem
Cross-vertical food, grocery, and dining intent is usually split across chatbots and brittle REST calls, so an agent can hallucinate orders or skip consent before money moves.
Approach
Nexus uses a Planner, Executor, and Synth loop: stage carts and table slots with read-only MCP tools, require human-in-the-loop approval, then run write tools. Calendar signals and quality-of-life context feed the same graph.
Architecture
Next.js Concierge UI streams SSE tool traces. FastAPI plus LangGraph StateGraph with HITL interrupts. Local mock MCP or a remote MCP endpoint. Independent proof of concept, not affiliated with Swiggy.