Cloud Infrastruct

From prompt to production. The Multi-Cloud AI agent that understands your vision in plain English and automatically builds your complete infrastructure.

Comment

Inspiration

The cloud is powerful, but deploying infrastructure is notoriously complex, slow, and requires deep expertise in tools like Terraform. This complexity creates a significant barrier, slowing down developers and making the cloud less accessible for students, startups, and managers. Some corporations worldwide have added further complexity by adopting the multi-cloud approach in order to gain flexibility , enhance reliability and resilience. We asked ourselves: What if you could bypass the complexity entirely? What if you could build your cloud infrastructure just by describing what you need in plain English? Our inspiration was to democratize cloud deployment, transforming it from a specialized, code-heavy task into a simple, conversational experience.

What it does

Cloud Infrastruct is a Multi-Cloud AI agent that acts as an autonomous cloud solutions architect, turning natural language prompts into fully deployed GCP / AZURE / AWS infrastructure. A user simply describes their goal, like "I need a scalable web server for my new application," and clicks send. From there, Cloud Infrastruct handles the entire lifecycle:

  • Plans an optimal architecture by reasoning about the user's request using Google Gemini and ElasticSearch RAG.

  • Provides an Architecture Diagram for the plan and the estimated monthly cost to run the plan.

  • Generates secure, pre-validated Infrastructure as Code (Terraform templates) based on the plan.

  • Can translate the the same architecture plan to other cloud providers and provide an architecture diagram and rough cost estimate.

It's a true multi-cloud "prompt-to-production" experience that makes cloud engineering as easy as having a conversation.

Key Features

  1. AI-Powered Architecture Planning: Uses Gemini-2.5-flash to analyze requirements Generates detailed technical specifications and architecture plans Provides cloud best practices and solution patterns for the target platform (GCP/ AWS / AZURE) via ElasticSearch RAG Considers performance, security, and cost optimization by combining the power of Gemini and ElasticSearch
  2. Visual Architecture Diagramming: Creates detailed text diagrams (ASCII)
  3. Intelligent Cost Estimation: Provides detailed monthly pricing breakdowns Creates formatted cost tables by service Considers usage patterns and scaling requirements Helps with budget planning and optimization
  4. Infrastructure as Code Generation: Generates Terraform IaC templates Includes proper resource tagging and naming Template validation and syntax checking
  5. User-friendly chatbot interface: Based on Google Agent Development Kit (ADK) Familiar chat interface similiar to Gemini and ChatGPT Text and voice support
  6. Multi-Cloud Support: Supports GCP, AWS and Azure.

How we built it

Project Genesis & Vision

We set out to build an AI-powered multi-cloud infrastructure assistant that could take natural language descriptions and turn them into production-ready cloud deployments. The vision was ambitious: create a system that acts like a enterprise multi-cloud engineer, handling everything from planning to deployment. We utilized the Google Agent Development Kit (ADK) to build cloud infrastruct and allow for easy integration between google services and elasticsearch.

Architecture Diagram

Infrastruct Architecture Diagram

Core Architecture Decisions

AI Agent Framework Selection

We chose Google ADK Framework for multi-agent orchestration

from google.adk.agents import LlmAgent
from google.adk.agents.sequential_agent import SequentialAgent
from google.adk.tools import AgentTool
from pydantic import BaseModel, Field

This gave us:

  • Tool chaining capabilities
  • Google Gemini and ElasticSearch integration
  • Structured agent workflows

Why Google ADK?

  • Out of the Box chat interface via adk web
  • Tool-based agent architecture
  • Easy to extend with custom tools
  • Handles complex multi-step workflows

Retrieval Augmented Generation (RAG) System

We chose ElasticSearch for our RAG system because its a powerful and flexible search engine with the following features:

  • Excellent Search Capabilities: Elasticsearch is built for finding relevant information quickly within large amounts of text. This is exactly what our PlannerAgent needs to do – search through your ingested best practice documents to find the guidelines relevant to the user's request.

  • Supports Hybrid Search: It handles both traditional keyword search (finding exact terms like "Cloud SQL" or "VPC") and semantic search (understanding the meaning behind the user's request, like "secure database," and finding conceptually similar text using vector embeddings generated by models like ELSER). This combination is crucial for RAG to retrieve the most relevant context.

  • Scalability: Elasticsearch is designed to handle large datasets and can scale out as you add more documentation or knowledge.

  • Data Ingestion and Processing: It provides tools like ingest pipelines that allow us to automatically process documents as they are added. We used this to run the ELSER model and create the searchable content_vector field during our setup.

  • Integration: It has mature client libraries , making it relatively straightforward to integrate into our ADK agent's tools .

Project Structure

We built specialized tools for each workflow step:

 app/

├── agent.py      # agent

├── ascii_diagrammer.py    # diagram helper  

├── tools.py   # agent tools

 docs/

      gcp/

      aws/

      azure/

├── .env     # environment file

├── ingest.py  # docs ingestion into elasticsearch script

├── requirements.txt  

└── run.py

Development Phases

Phase 1: Data Gathering and Sorting

We gathered pdf document data on the three cloud providers (GCP, AWS ,AZURE) looking for:

  • Well Architected Framework Guidelines for the particular cloud provider
  • Security best practices
  • Service offerings descriptions
  • sample architectures by use case

The pdfs where then sorted into their corresponding folders


 docs/

      gcp/

      aws/

      azure/

Phase 2: Data Ingestion into ElasticSearch A serverless project was created in ElasticSearch Console then manual indexes created under that projects that corresponded to the folders labels under the local docs folder. The endpoint url and the api key where obtained then added to the .env file. An ingest.py script was created to:

  • Define the Ingest Pipeline to run ELSER.
  • Create the index mappings (to store the vectors).
  • Use LangChain to chunk your PDFs.
  • Bulk-upload the chunks to the correct index.

After running the ingest script, our documents where uploaded and ready for searching. The ingest script only needs to be executed once.

Phase 3: Chatbot creation using Google ADK and ElasticSearch Integration

For our use case, the multi-agent sequential design pattern was the best fit. We needed a logical chained flow from generating the cloud architecture plan to generating the architecture diagram to determining the estimate cost to generating the terraform Iac template and finally presenting the result to the user.

Infrastruct Architecture Diagram

Benefits:

  • Reduced latency
  • Reduced operational costs

Once the agent was built with Gemini as the brains, we built the tool that would query ElasticSearch based on target platform (GCP, AZURE, AWS) on best practices to provide context to our agent's Gemini brain. We finalized by creating sub agents for diagramming, cost estimation and IaC generation

Development Tools & Technologies

Core Stack:

  • Python 3.13.6: Main development language
  • Google ADK: Agent orchestration and Chatbot web-based User Interface
  • ElasticSearch: Retrieval Augmented Generation
  • Google AI Studio: AI model hosting

AI & Automation:

  • Gemini-2.5-flash: Primary AI model

Visualization:

  • ASCII Art: Immediate diagram display

Challenges we ran into

We ran into some issues with the latest ADK installation we have to downgrade to another version.

Accomplishments that we're proud of

We successfully built a comprehensive multi-cloud AI-powered infrastructure assistant that:

  • Reduces cloud engineering time from hours to minutes
  • Builds cloud infrastructure considering platform best practices
  • Provides learning opportunities through detailed explanations
  • Scales from simple to complex infrastructure needs
  • Fully terraform compatible

The project demonstrates how AI agents and ElasticSearch RAG can transform complex technical workflows into accessible experiences.

What we learned

The ElasticSearch platform has powerful and perfomant search capabilities that meet the workload demands of modern applications. The cloud infrastruct project had multi-cloud RAG capabilities thanks to the search and indexing technology offered by ElasticSearch.

What's next for Cloud Infrastruct

Allow for direct deployment of generated architecture plans to cloud via terraform enterprise API

Built With

Share this project:

Updates