Handbook for Agents IRL Hackathon

Welcome to the Agents IRL! This handbook provides you with essential resources, tools, and tips to build and deploy your agentic AI applications using Google Cloud Run with serverless NVIDIA L4 GPUs.

Prerequisites:

Please have a gmail account ready before the Hackathon so you can sign in to the Google Cloud project to redeem the free GPU credits during the event.

Google Cloud Run Handbook

For more resources for utilizing Google Cloud Run, the Cloud Run team has provided a handbook here with examples for you to review

Technical Resources

• Google Cloud Run Serverless GPUs
• Cloud Run Overview
• Using NVIDIA L4 GPUs on Cloud Run
• Each team gets 1 GPU per Google Cloud project.

TEAM formation requirements: max 5 ppl and min 2 ppl. per team

• Open Model Examples
• Gemma 3 (1B/4B/12B)
• DeepSeek-r1 7B
• Llama 2 7B

• Stable Diffusion XL (SDXL)

• Model Access & APIs
• Documentation links and example code snippets, please refer to the Google Doc here
• Use lightweight inference optimized for serverless environments.

Project Ideas

• Autonomous agents that perform multi-step tasks using Llama 2 or Gemma 3.
• AI-powered search or recommendation systems using DeepSeek-r1.
• Creative generation apps leveraging Stable Diffusion XL for image synthesis.
• Agentic workflows combining multiple models for complex decision-making.
• Real-time AI assistants deployed on Cloud Run with GPU acceleration.

Best Practices

• Team Formation: Lock in your team early during the Day 1 session.
• Development: Use Cloud Run’s scalability to test and iterate quickly.
• Resource Management: Optimize GPU usage to stay within limits.
• Debugging & Monitoring: Use Opik by Comet to track model performance and safety.
• Demo Preparation: Prepare a clear, concise live demo showcasing your app’s agentic capabilities.
• Submission: Submit your project by 10:00 AM on Day 2 for judging.

Good luck, have fun building, and don’t hesitate to ask mentors for help during the event!