Nexus

Inspiration

During emergencies, critical information is often scattered across weather systems, maps, incident reports, resource databases, and human observations. The challenge isn't simply collecting information—it's turning that information into clear, timely decisions.

This inspired me to build NEXUS, a multi-agent AI decision-support system designed to help emergency operators respond to urban flooding faster and more effectively.

What I Built

NEXUS follows the workflow:

Incident → Investigate → Assess → Allocate → Simulate → Human Approval → Monitor → Re-plan

I built specialized AI agents for:

Coordinator — breaks an incident into operational tasks and orchestrates the system. Research — gathers relevant crisis information. Risk — produces explainable risk assessments. Geospatial — analyzes affected zones, roads, shelters, and critical facilities. Resource — allocates teams, vehicles, and emergency supplies. Simulation — evaluates "what-if" scenarios. Decision — generates actionable response strategies for human approval.

One of the key features is the simulation engine. For example, if a major road becomes inaccessible, NEXUS recalculates routes, response times, resource allocation, and risk, then generates an updated recommendation.

I also built a multimodal workflow where an image of road damage can be analyzed by Gemini, converted into a structured incident, and used to update the operational state and recommendations.

How I Built It

I used Google Cloud technologies throughout the system:

Google ADK for multi-agent orchestration Vertex AI / Gemini for reasoning and multimodal analysis Cloud Run for the backend and application services Firestore for operational state Pub/Sub for event-driven updates Cloud Storage for uploaded documents and images

To make the demo reliable and reproducible, I created a deterministic synthetic urban-flooding dataset containing zones, roads, shelters, hospitals, response teams, vehicles, supplies, weather events, and incident reports.

A key design principle was separating computation from AI reasoning. Deterministic code handles operational calculations such as resource availability, routes, capacities, and risk inputs. Gemini focuses on reasoning, synthesis, explanations, and planning.

What I Learned

Building NEXUS taught me that effective AI systems aren't necessarily about giving one model responsibility for everything. A better approach can be to combine specialized agents with deterministic tools, allowing each component to focus on what it does best.

I also learned the importance of human-in-the-loop decision making. NEXUS doesn't autonomously make consequential emergency decisions. Instead, it presents recommendations, trade-offs, confidence levels, and simulations so that a human operator can make the final decision.

Challenges I Faced

As a solo developer, one of my biggest challenges was designing a system that felt powerful without becoming unnecessarily complex.

I initially considered creating more specialized agents, but I learned that adding agents simply for the sake of having more agents can make a system harder to understand and maintain. I therefore kept the architecture focused on the agents that provide meaningful value.

Another challenge was preventing AI from inventing critical operational numbers. I addressed this by using structured data and deterministic calculations for important values such as resource availability, route constraints, shelter capacity, and risk inputs.

I also had to make a complex multi-agent system understandable within a short demo. I solved this through a command-center interface with a live map, agent activity timeline, recommendation cards, and a "What if?" simulation workflow.

Why NEXUS?

NEXUS is not designed to replace emergency operators.

It is designed to give them a clearer understanding of what is happening, show what could happen next, explain the available options, and help them make better decisions faster.

NEXUS turns fragmented crisis information into actionable, explainable, and adaptive response plans—while keeping humans in control.