MallCore-Agent

An autonomous MCP-powered AI agent using MongoDB Atlas and Gemini to orchestrate real-world mall logistics and beyond-chat automation.

Comment

Inspiration

Traditional chatbots are limited to answering questions within the confines of a browser window. In the context of brick-and-mortar retail and large shopping malls, customers face fragmented experiences: checking stock on one app, making restaurant reservations on another, and guessing where to park. We were inspired to bridge the gap between digital AI reasoning and physical world operations by creating an autonomous agent that acts as a real-time logistics coordinator, turning natural language intent into verified multi-step physical actions.

What it does

MallCore-Agent is an autonomous logistics ecosystem designed for physical malls. When a user inputs a complex, multi-step mission (e.g., "I need to buy a handmade greeting card, have a traditional dinner, reserve a table automatically, and know where to park nearby"), the agent:

  1. Plans & Reasons: Analyzes the user's ultimate goals and coordinates store locations, real-time inventories, restaurant availability, and parking structures.
  2. Executes Beyond-Chat Actions: Automatically writes to the live database to book a reservation, securely updating availability metrics without human friction.
  3. Consolidates an Itineray: Grounded entirely in real-time data, it delivers an optimal, location-aware route, telling the user where to shop, confirming their dinner slot, and recommending the exact parking zone to minimize walking distance. ## How we built it The core architecture is built upon an integrated Model Context Protocol (MCP) server architecture to grant Gemini structural superpowers over live operational data:
  4. The Brain: Powered by the advanced reasoning capabilities of Gemini 2.5 Flash to break down high-level multi-step missions into sequential tool calls.
  5. The Engine: An integrated backend written natively in Node.js and Express, utilizing the official @modelcontextprotocol/sdk to establish seamless Server-Sent Events (SSE) data streams.
  6. The Superpowers (Data Grounding): Orchestrated dynamically through a cloud cluster on MongoDB Atlas. We leveraged three custom collections—stores_inventory for product availability tracking, restaurants_venues for live capacity management, and mall_map_nodes for intelligent spatial mapping. ## Challenges we ran into One of our biggest hurdles was navigating rigid environment constraints and fast-evolving SDK versions during rapid prototyping. We initially hit environment validation dead ends when deploying complex agent infrastructure wrappers across operating systems. Realizing the strict timeline, we pivoted towards building a highly resilient, streamlined architecture directly within our Node.js runtime environment. This shift allowed us to bypass local Python wrapper conflicts and execute native, bulletproof database transactions directly through our MCP payload loops. ## Accomplishments that we're proud of
  7. True Beyond-Chat Mechanics: We built a loop where the agent doesn't just display database context—it interacts with it securely, accurately executing live updates ($inc: { available_tables_7pm: -1 }) automatically in Atlas.
  8. Context Isolation & Grounding: Successfully preventing hallucinations by locking the model’s reasoning strictly to the live state of our active MongoDB collections.
  9. Architectural Agility: Adapting our architecture under tight hackathon pressure to deliver a robust, functional, and fully integrated production file before the deadline. ## What we learned We gained deep insights into the power of the Model Context Protocol (MCP) as the standardized connective tissue for the next generation of AI. We learned that by combining Gemini's raw reasoning faculties with specific, transactional database constraints, we can build highly secure, self-correcting agent systems capable of transforming real-world physical workflows. ## What's next for MallCore-Agent We plan to scale MallCore-Agent into a fully distributed multi-agent system. Future iterations will include real-time vehicle telemetry hooks to dynamically update free parking slots using MongoDB geospatial queries ($geoWithin), automated tenant multi-tenant campaign triggering, and direct integration with live payment gateways for frictionless physical-to-digital checkout experiences.
Share this project:

Updates