Inspiration
Brick-and-mortar retail environments are massive, fast-moving ecosystems. When a storefront hazard or facility emergency occurs at a mega-mall, manual communication delays between tenants, operations managers, and field technicians can result in severe property damage or lost revenue. We wanted to move beyond typical "Q&A chatbots" and build an intelligent, autonomous operator capable of instantly assessing facility risks against actual compliance data, making live state evaluations, and directly updating backend operational databases without human delay.
What it does
MallOps Intellect is an autonomous, multi-step operations manager built explicitly for facility maintenance and tenant compliance workflows. When a user or tenant reports an issue (e.g., an electrical malfunction or a water leak), the agent executes a structured 5-step mission:
- Grounding Evaluation: It queries a custom Vertex AI Search data store containing mall compliance policies to see if the incident mandates immediate utility isolation (like an immediate power or water shutdown).
- Database Tenant Validation: It queries an external MongoDB database using Model Context Protocol (MCP) tool extensions to verify the tenant's lease validity and occupancy records.
- State Evaluation: It logically processes the policy rules against the tenant records to dynamically calculate the correct ticket priority (Low, Medium, High).
- Database Write Automation: It executes an MCP tool call to create and log a structured maintenance ticket directly into the live database.
- Human-in-the-Loop Verification: It presents the recorded payload object back to the operator for final safety confirmation.
How we built it
We constructed the core agent flow within the Google Cloud Agent Platform Console, leveraging Gemini 3.5 Flash as our primary model for fast reasoning and swift demonstration latency.
- Data Grounding: We utilized unstructured text files parsed through Vertex AI Search data stores to keep the agent tightly tethered to true compliance manuals, effectively eliminating hallucinations.
- Partner Power (MCP): We built a public OpenAPI mock server endpoint running on Railway (
https://express-post-production.up.railway.app/mcp) to simulate a live MongoDB database schema, giving the Gemini model real-time "superpowers" to seamlessly run database actions likequery_tenantandcreate_ticket.
Challenges we ran into
One of our biggest hurdles was keeping an advanced LLM focused on a rigid, sequential execution order. In early iterations, the model would attempt to write to the database before verifying the tenant's legal status or policy constraints. We solved this by structuring a definitive System Execution Flow layout right inside the agent instructions, transforming its behavior into a deterministic step-by-step state machine. We also had to navigate Google Cloud's IAM permission structures to cleanly map our agent workspace to our Vertex AI Search data assets.
Accomplishments that we're proud of
We are proud to have built a solution that completely moves away from basic conversational chat and behaves as a genuine operational utility. Successfully connecting external OpenAPI tool specs via the Model Context Protocol allows the agent to handle real-world write actions natively, transforming a standard large language model into an active, hands-on enterprise coordinator.
What we learned
We learned that giving an AI agent specific "muscles" (database access via MCP) and clear "memories" (grounded documents) vastly diminishes the typical predictability risks associated with generative AI. Using Gemini 3.5 Flash showed us how fast an agent can evaluate multi-variable logistics on the fly without sacrificing system latency.
What's next for MallOps Intellect
We plan to upgrade the server endpoint to a production-grade MongoDB instance requiring secure API authentication keys. Additionally, we would love to connect the agent directly to real-time IoT facility hardware triggers (such as smart water meters or electrical grids) so that MallOps Intellect can automatically generate its own tickets before a human tenant even notices a malfunction.
Log in or sign up for Devpost to join the conversation.