MudançAI

MudançAI is an interactive moving assistant and agentic orchestration system for planning stress-free residential moves. It turns the chaotic process of packing into a structured, measurable experience: agents convert unstructured user input into a structured Inventory, generate Packing Plans validated against logistics and fragility rules, and securely export this data to productivity tools.

The first visible surface is a chat-based Generative UI experience, but the product is not just a standard Q&A chatbot. The chat interface is the testbed for a larger cognitive loop: dynamically parse messy user data (text descriptions or room photos), categorize the risk level and volume of each item, test the best distribution configuration across boxes, and maintain this complex state over weeks of preparation.

MudançAI is designed around a secure orchestration model using LangGraph and CopilotKit. The agent does not take destructive actions or generate arbitrary frontend code at runtime. It produces declarative state updates based on pre-approved Tools, while the client renders the progress through a trusted BFF (Backend-for-Frontend) layer. The system aligns this model with Human-in-the-Loop (HITL) workflows, ensuring the agent requests user approval before finalizing boxes or exporting data via the Model Context Protocol (MCP).

The project does four things:

Generate Dynamic Inventories: Given a simple description or room image, the system identifies, catalogs, and classifies items (e.g., fragile, heavy, electronic) using the Gemini API.
Create Intelligent Packing Plans: A reasoning engine validates and distributes the cataloged items into optimized virtual “Boxes,” ensuring heavy items don’t crush fragile ones and preventing unnecessary mixing of belongings from different rooms.
Maintain Context (Durable Threads): Through an Intelligence layer powered by Postgres and Redis, the system saves the progress. A user can start organizing the living room on Monday, resume the kitchen on Friday, and never lose the history of what has already been packed.
Sync with the Real World (MCP): Validated tools export the final plan as actionable tables and checklists directly into the user’s workspace (such as Notion), making the plan ready to execute on moving day.
The north-star product is an AI-powered moving control tower. Homeowners should get a foolproof, stress-reducing packing guide and checklist. Moving Companies (in the future) should get accurate volume estimates and fragility reports before arriving on site. Developers should get a robust template for combining LangGraph, Next.js, and MCP servers into a real-world, task-oriented agent architecture.

Prior Work & Hackathon Contributions

Prior Work (What we started with): Our project is built upon a CopilotKit Agent Template (originally the v2a-notion-lead-form template). This provided us with a solid scaffolding and boilerplate architecture, which included:

The base Next.js frontend with the CopilotKit Generative UI chat interface.
The BFF (Backend-for-Frontend) layer running on Hono, which proxies requests to the agent.
The foundational LangGraph setup for running an agent.
The base CopilotKit Intelligence infrastructure (Postgres/Redis) for durable thread persistence.
The boilerplate integration for the Notion Model Context Protocol (MCP) server.
Created During the Hackathon (What we built): During the hackathon, we completely transformed this lead-generation template into MudançAI, a specialized cognitive agent for residential moving. Our specific contributions include:

Core Domain Logic & Prompts: We wrote entirely new system prompts and interaction policies for the agent, shifting its persona from a simple form-filler to a proactive moving assistant.
Dynamic Inventory Parsing (parse_inventory): We integrated the Gemini API to take unstructured user input (text or images about their rooms) and algorithmically categorize them into structured Items (assessing weight, fragility, and room origin).
Packing Engine (generate_packing_plan): We developed a custom reasoning tool that structures the parsed inventory into optimized, virtual Boxes. The logic prevents heavy items from crushing fragile ones and organizes boxes logically by room.
State Schema Updates: We overhauled the LangGraph state schemas and data models to support our new Items and Caixas (Boxes) structures, replacing the old lead-form schemas.
Workflow Orchestration: We implemented a Human-in-the-Loop (HITL) transition policy, ensuring the agent correctly requests user approval for the packing plan before using the Notion MCP to export the final checklist.
Rebranding & UX Refinement: We adapted the frontend components and backend tool names to align with the “MudançAI” brand and the moving use case.
In short, while we leveraged CopilotKit’s excellent architecture for the plumbing (chat UI, thread persistence, and MCP connection), all of the moving-specific AI logic, data schemas, packing algorithms, and Gemini integrations were built from scratch during the hackathon.