Agent Krimini

An AI agentic tool that helps you minimize the risk of dangerous incidents on campus.

Comment

Inspiration

Campus safety has been a big concern for high school and university students, so I want to build this as a software for campus police, crime analysts, and students to improve campus protection and reduce crime risks at campuses.

What it does

Agent Krimini utilizes a sophisticated Multi-Model Orchestration strategy, leveraging different tiers of the Gemini 3 and 2.5 series based on the complexity and latency requirements of each task:

1. The Strategic Brain: gemini-3-pro-preview

  • Role: Reasoning & Safety Analysis.

  • Application: Used in the Safety Dashboard to generate the overall campus safety score and tactical recommendations.

  • Why: It utilizes a thinking Budget of 2,000 tokens to perform deep reasoning across multiple concurrent incidents before providing its final safety assessment.

2. The Native Voice Interface (gemini-2.5-flash-native-audio-preview-12-2025):

  • Role: Real-Time Multimodal Interaction.

  • Application: Powers the Voice Link feature.

  • Why: This model is designed specifically for low-latency, "native" audio-to-audio conversations. It processes raw PCM audio streams directly, allowing for human-like speed and interruptions without the delay of traditional speech-to-text-to-speech pipelines.

3. The Tactical Workhorse (gemini-3-flash-preview):

  • Role: Structured Data & Speed.

  • Application: Handles Incident Analysis (JSON classification) and Emergency SOS Drafting.

  • Why: It is optimized for high-speed performance and strict adherence to JSON schemas, ensuring that incident data is categorized correctly for the NoSQL database instantly.

4. The Safety Chat Assistant: gemini-2.5-flash

  • Role: Grounded Reasoning.

  • Application: Powers the Strategic Reasoning Chat.

  • Why: It is used here because of its robust support for Google Maps Grounding, allowing the agent to "see" campus locations and provide real-time spatial links to the user.

5. Improve Overall Safety by 80% with instant crime reports and updates.

  • Discrete, safe for reporting and preventing any incidents related to students.

  • Estimate the safety of the students with alternative evacuation at safe locations.

How we built it

I built a multi-agent orchestration system in Google AI Studio that dynamically routes tasks to specialized Gemini models based on reasoning complexity, latency requirements, and modality. This program operates using TypeScript and React.js, and integrates multiple Gemini models into a coordinated pipeline that supports incident analysis, spatial reasoning, voice interaction, and strategic safety assessment in a unified experience. The website was deployed to Google Cloud Run.

The Safety Dashboard utilized gemini-3-pro-preview to provide the overall campus safety score and tactical recommendations.

The gemini-2.5-flash-native-audio-preview-12-2025 model were use for the AI Voice Agent feature to safely and quickly interact with AI Agents for safe procedures. I've also implemented the Gemini 2.5 Flash Model for the Safety Chatbot Assistant with core reasoning and safety navigation.

Challenges I ran into

The only problem I had with this application was the implementation of the Agent Development Kit to enable different functionality in the build. I'm also having some logic conflicts with the ideas to implement surveillance systems for crime detection, but this will be in future development.

Accomplishments that I'm proud of

I'm proud to completed this project using the Gemini 3 Pro Model, and we're able to have a working prototype that specifically aims to increase campus safety with the ability to pinpoint any dangerous situations and the ability for AI Agents to navigate the students to safety.

What I learned

I learned that multimodal AI becomes far more powerful when models are specialized and coordinated. Using the right model for the right task dramatically improves responsiveness, accuracy, and usability.

What's next for Agent Krimini

In the future, I plan to expand real-time data integration, improve predictive risk modeling, and explore partnerships with high school and university campuses to transform Agent Krimini from a prototype into a scalable, real-world safety platform. I would love to add other technologies to improve visibility and detailed navigation.

Built With

  • gemini-2.5-flash-native-audio
  • gemini-3-pro
  • google-ai-studio
  • google-cloud-run
  • html
  • nosql
  • react
  • typescript
Share this project:

Updates