AImpact

We built AImpact with the intention of automating marketing tasks and basically having our agent work alongside business developers and digital marketers. We have around 200 n8n automation workflows

Comment

Inspiration

The inspiration for AImpact came from the potential of AI agents to streamline complex workflows, particularly in automating repetitive tasks across multiple tools. As a developer accustomed to traditional application development, I was fascinated by the Google Cloud Multi-Agent Hackathon’s challenge to build collaborative, multi-agent systems. My goal was to create an AI agent that could orchestrate tasks using external tools, reducing manual effort in real-world scenarios like data processing or customer support automation. I aimed to integrate Google’s Agent Development Kit (ADK) with n8n, a powerful no-code workflow tool, to demonstrate how AI can enhance workflow automation.

What I I Learned

Building AImpact was my first venture into AI agent development, and it was a steep learning curve. I learned:

  • ADK Framework: How I structured my code to define agents and tools using Google’s ADK, leveraging Gemini models for task processing.
  • n8n Integration: I mastered calling n8n workflows via API, passing dynamic parameters, and handling responses.
  • Vertex AI Agent Engine: Deploying a scalable runtime on Google Cloud was new to me; I learned to configure dependencies, environment variables, and manage permissions.
  • Cloud Deployment: I explored Google Cloud services like Cloud Storage and Cloud Run, gaining hands-on experience with cloud-native development.
  • Rapid Prototyping: Using adk-web allowed me to focus on agent logic rather than building a custom frontend, a key lesson in prioritizing functionality for hackathons.

How I I Built It

AImpact is an AI agent built with Google’s ADK*, powered by a Gemini model, and deployed on **Vertex AI Agent Engine. The agent interacts with users through adk-web, a built-in ADK tool for testing interface. It processes user requests and triggers actions by calling **n8n* workflows via API, which execute tasks like data transformation or external API calls. The project architecture includes:

  • Agent Logic: Written in Python, using ADK to define the agent and a custom tool for n8n integration.
  • n8n Workflows: Hosted on a self-hosted n8n instance (or n8n cloud), configured to handle specific tasks.
  • Deployment: Hosted on Vertex AI Agent Engine for scalability, with adk-web deployed on Render (or Cloud Run) for a public UI.
  • Dependencies: Managed via requirements.txt, including google-cloud-aiplatform, requests, and pydantic.

Challenges I I Faced

As a first-time AI agent developer, I encountered several challenges:

  • n8n Integration: Ensuring reliable API calls from the agent to n8n required debugging network issues and handling error responses.
  • Vertex AI Deployment: Configuring environment variables and permissions for the Vertex AI service account was complex, especially for a cloud newbie.
  • Time Constraints: Balancing agent logic with deployment setup under the hackathon deadline was tough. I chose adk-web over a custom frontend to save time.
  • Learning Curve: Grasping ADK’s concepts, like agent orchestration and tool definition, took significant effort, but Google’s documentation and samples helped.

Despite these hurdles, AImpact demonstrates the power of AI-driven automation, and I’m excited to share my journey with the hackathon community!

What it does

  • Autonomous Tool Execution : It can automatically execute specialized tools for tasks such as SEO keyword generation, lead nurturing, converting YouTube videos to Twitter threads, and scraping Reddit for content ideas. It infers user intent and uses the appropriate tool with minimal input.
  • Intelligent Input Handling : The agent is designed to normalize and repackage user inputs to fit the schemas of its tools. It can convert single strings to lists, join lists into strings, and infer missing information from context or use sensible defaults.
  • Error Resilience : If a tool encounters an error, the agent attempts to fix inputs, retry the operation, or fall back to a simplified execution. It only informs the user if an issue is unresolvable, providing clear explanations and next steps.
  • Polished Output Delivery : The agent presents results in a clear, structured format. For tool outputs, it returns the raw JSON or string result directly, unless specific formatting is requested. It also suggests next steps and offers options for exporting data.
  • Proactivity and Scalability : It anticipates user needs by suggesting related tools and seamlessly integrates new tools by matching their capabilities to user intents. It can also sequence tools logically to achieve comprehensive results.
  • User-Centric Efficiency : The agent minimizes user effort by avoiding unnecessary questions and maintaining a concise, professional, and engaging tone to provide a premium user experience. In essence, the AImpact project provides an advanced AI assistant that streamlines marketing and business intelligence operations by intelligently automating tasks and delivering data-driven outputs.

Accomplishments that we're proud of

  • Robust Error Handling and Debugging : We effectively identified and resolved critical deployment issues, including FileNotFoundError related to extra_packages and ModuleNotFoundError for the agents module. This involved meticulous debugging and understanding of Vertex AI's agent deployment requirements.
  • Dependency Management : We successfully managed and resolved missing Python dependencies like pydantic and cloudpickle during the deployment process, ensuring all necessary libraries were included for the agent's operation.
  • Integration of Specialized Tools : The agent successfully integrates and orchestrates various specialized tools for marketing and business intelligence, such as SEO keyword generation, lead nurturing, YouTube to Twitter thread generation, and Reddit content idea scraping.
  • Autonomous and Intelligent Agent Design : The AImpactSuperAgent is designed with core principles of autonomous tool execution, intelligent input handling, error resilience, and polished output delivery, making it a highly effective and user-centric AI assistant.
  • Scalability and Proactivity : The architecture allows for seamless integration of new tools and proactive suggestions to users, enhancing the agent's utility and adaptability.

What's next for AImpact

Now we have the main orchestrator agent in place, we plan to add over 200 pre built workflows and transfor AImpact into a true Super Agent

Built With

  • adk
  • adk-web:
  • ai
  • api.
  • cloud
  • cloudpickle
  • fastapi
  • gcs
  • gemini-1.5-pro).
  • generetiveai
  • google
  • google-auth
  • n8n:
  • pydantic
  • python-dotenv
  • requests
  • storage:
  • uvicorn
  • vertex
Share this project:

Updates