Skip to main content

2 posts tagged with "development"

View All Tags

Build Your Own Kubernetes Co-Pilot: Harness AI for Reliable Cluster Management

ยท 8 min read


Have you ever felt frustrated by nonsensical AI outputs and hallucinations? If yes, this blog is going to be helpful for new or seasoned Kubernetes users who want to explore how AI can help manage Kubernetes resources more reliability.

What are AI hallucinations?โ€‹

In a nutshell, AI hallucination occurs when a large language model (LLM) generates misleading or incorrect information in response to a prompt. This can happen due to various factors such as insufficient or flawed training data, overfitting, unrecognized idioms or slang, and adversarial inputs. These hallucinations manifest when the AI, aiming to produce coherent responses, makes errors that range from subtle factual inaccuracies to nonsensical or surreal outputs, similar to how humans might perceive patterns in random visuals.

In the context of Kubernetes, these aren't just minor nuisances; they can lead to significant operational blunders. In this blog, we explore how to enhance reliability of AI responses, mitigate the risks of hallucinations, manage Kubernetes resources using AI!

How can AI be helpful in managing Kubernetes resources?โ€‹

Before we start exploring the technical setup, let's answer the question how can AI be helpful in managing Kubernetes resources? Imagine an AI assistant that can help you create, fix, and validate Kubernetes resources in a conversational manner. You might ask it to create a new deployment, fix a broken service, or validate a YAML file. If you are learning Kubernetes, this assistant can be a great learning tool to help you explore the cluster and clarify Kubernetes concepts.

Kubernetes helps manage cloud applications, but its YAML configurations can be tricky. When working with AI tooling, we've all faced those moments when AI tools, designed to ease this burden, instead contribute to it by generating nonsensical outputs; a phenomenon we refer to as "AI hallucinations".

Problem Statementโ€‹

Let's state the issue we do have with AI in the context of Kubernetes:

  • ๐Ÿค– AI faces issues with consistency and reliability when dealing with large YAML files.
  • ๐Ÿง  AIs can have "hallucinations," generating illogical outputs that become more problematic as the input size increases.
  • ๐Ÿ“ˆ This inconsistency makes working with AI models non-deterministic and error prone


Our main goal is to increase reliability and consistency in AI responses. We use two main techniques to achieve this:

  • ๐Ÿ› ๏ธ Function calling to bind API routes as tools available for the AI Assistant to communicate with a Kubernetes cluster
  • ๐Ÿ” Internet search APIs to provide accurate and relevant information about Kubernetes

Implementation Planโ€‹

The following steps outline the plan to achieve our goals:

  • ๐Ÿ’ผ Use Flowise to implement the logic flow so that the AI Assistant can help with managing and troubleshooting a Kubernetes cluster on our behalf.
  • ๐Ÿ› ๏ธ Create a simple Flask API that exposes functions for the AI Assistant to enable it to interact with the Kubernetes cluster.
  • ๐Ÿ’ป Use function calling to bind the API routes as tools available for the AI Assistant which enables communication with a local Kind cluster with Kubernetes running.
  • ๐Ÿ’ฌ Test the AI Assistant with various scenarios to ensure it can handle different Kubernetes configurations and provide accurate responses.

Assistant in Action

To follow along, you can clone the repository from GitHub, install prerequisites and follow the instructions.

Step 1: Setup the AI Assistantโ€‹

In flowise create a new assistant. Notice that I'm using OpenAI's latest model, but for testing purposes you can select less powerful models or any open source model. The quality of responses will be affected, but it will still work.

Here are instructions that the assistand will follow:

You are a helpful Kubernetes Assistant specializing in helping build, fixing and validating various kubernetes resources yaml files.
Start by greeting the user and introducing yourself as a helpful and friendly Kubernetes Assistant.

If the user asks for help with creating or validating yaml files, do the following:

- if the files are correct proceed with the next steps, if no propose fixes and correct the file yourself
- if user asks for information about the kubernetes cluster use the get_config function and provide relevant information
- ask the user to submit one yaml file at a time or create one yaml file yourself if the user asks you to create one
- send the YAML content and only the YAML content to the create_yaml function
- immediately after use the tool cleanup_events to clean any old events
- ask the user if they would like to see the validation results and inform them that it takes some time for the resources to be installed on the cluster
- if the user responds yes, use the tool check_events to see if everything is correct
- if the validation passes, ask the user if they want to submit another YAML file
- if the validation fails, propose a new corrected YAML to the user and ask if the user would like to submit it for validation
- repeat the whole process with new YAML files

Your secondary function is to assist the user in finding information related to crossplane. Example categories:

- for questions about kubernetes concepts such as pods, deployments, secrets, etc, use brave search API on
- for generic Kubernetes questions use brave search API on kubernetes docs:
- for questions regarding kubernetes releases and features use brave search API on kubernetes releases documentaiton: If you are asked for details about specific release, select one of the releases, otherwise use latest stable release.

Step 2: Flask APIโ€‹

The file defines API routes that wrap the kubectl commands.

โ„น๏ธ The flask server is a naive implementation for demonstration purposes only. In real life scenario, we wouldn't call kubectl directly from the server but rather use a client library like kubernetes or client-go.

Step 3: Expose local URL to the internetโ€‹

In order to enable the OpenAI assistant to use the functions we must expose the locally running flask server to the internet. For this a nice tool to use is ngrok. You can download it from here and follow the instructions to expose the local URL.

Step 4: Function callingโ€‹

Now we can create functions for each API route. Those are:

  • get_config - returns the current Kubernetes configuration
  • create_yaml - creates a new Kubernetes resource from a YAML file
  • check_events - checks the status of the Kubernetes resources

For each of those routes we create a function that calls the API and returns the response. Here is how the function looks like in flowise:


Step 5: Use brave search APIโ€‹

The secondary function of our assistant is to assist the user in finding information related to Kubernetes. We can use the brave search API to achieve this

Step 6: Testingโ€‹

Now since we have the whole flow available, let's test the assistant.


Let's start by asking what is the cluster we are running on:


Here the assistant used the get_config function to get the current Kubernetes configuration and correctly identified the cluster.

Now let's ask the assistant to create a new nginx based ingress:


Notice how the assistant correctly selected the create_yaml function to create the ingress and then used the check_events function after asking if we would like to see the output. It's also interesting that it has found a different event that was not related to the nginx ingress and classified it as unrelated to our request.

Now, let's submit a broken deployment and see if the assistant can fix it:


In this case we have submitted a broken deployment and the assistant has correctly identified the issue and even proposed a fix.

Lastly, let's check if the assistant can help us undrstand some Kubernetes concepts:


Here the assistant has used the brave search API to find information about the Kubernetes resource model and provided a link to the source.

Closing Thoughtsโ€‹

We have successfully demonstrated that using function calling and carefully crafted prompt instructions, we can increase the reliability and usefulness of AI assistants in managing Kubernetes resources. This approach can be further extended to other use cases and AI models.

Here are a few use cases where this approach can be useful:

  • ๐Ÿค– improved learning experience
  • ๐Ÿ“ˆ help increase Kubernetes adoption
  • ๐ŸŒ virtual Kubernetes assistant

This guide demonstrates using function calling and carefully crafted prompt instructions to enhance the reliability and usefulness of AI assistants in Kubernetes management. These strategies can be extended to other use cases and AI models

Next Stepsโ€‹

Give it a try, build your own AI powered Kubernetes management today:

  • Clone the Repository: Visit GitHub to get the necessary files.
  • Set Up Your Assistant: Follow the instructions setup prerequisites and start building your Kubernetes Co-Pilot.
  • Engage with the Community: Share your experiences and solutions, there setup is very much proof of concept and can be improved in many ways.

Thanks for taking the time to read this post. I hope you found it interesting and informative.

๐Ÿ”— Connect with me on LinkedIn

๐ŸŒ Visit my blogs on Medium

Development with AI: the GAG Stack

ยท 5 min read



Are developers going to be replaced by AI? What is the future of software development? Those questions are asked again and again as the software development landscape is evolving rapidly.

Viewpoints are polarized and generate heated debates and discussions. There is enough debate to fill a book, but in this article, I would like to explore practical applications of AI in software development. We are operating under the assumption that AI is here to stay and evolve, but at the end of the day, it is a tool that can be used to enhance our capabilities.

The GAG Stackโ€‹

The GAG Stack is a bit of a tongue-in-cheek term that I came up with to describe a workflow that I have been experimenting with. It stands for GPT Pilot, Aider, and GitHub Copilot. These are three AI tools that exemplify well the stages of software development.

Communication and collaboration between people is at the heart of software development. For as long as this stays the case, AI tools will be used to help us model this process. This is how it could look like using the GAG Stack:


We will still have to gather requiremetns, design, refine, test, retest, fix bugs, debug and deploy. The paradigm doesn't change much, the tools however do. The tools evolved to help us with the process.

Example Workflowโ€‹

Let's take a look at how the GAG Stack could be used in practice. We will use a simple example of building a to-do list app.

Setup the environmentโ€‹

I'm using neovim and linux for my development workflow, yours might be different. Refer to the installation instructions for all the tools to setup on your machine.

For me the gtp-pilot runs via docker-compose, aider is installed via pip and GitHub Copilot as a neovim plugin.

Design and Refinementโ€‹

We start by gathering requirements for our to-do list app. We want to have a simple app that allows us to add, remove and edit tasks. Let's start by providing this concept to GPT Pilot.

The Docker image has only node installed, so we are going to use it. It should be simple to add new tools to the image or use local setup. Here is initial prompt for a simple todo app:


The main value of this tool is the ability to refine and iterate on the desing. GPT Pilot will ask for specifications and generate an initial scaffolding:


As a result of this back and forth, GTP Pilot generated app in a local folder (mounted via volume in docker-compose):

~/gpt-pilot-workspace/minimal-todo-app๐Ÿ”’ [๎œ˜ v16.15.1]
โžœ tree -L 3 -I node_modules
โ”œโ”€โ”€ app.js
โ”œโ”€โ”€ package.json
โ””โ”€โ”€ package-lock.json

0 directories, 3 files

After a few iterations, we have a simple app running:

app-running with the following code:

// Require Express and Body-parser modules
const express = require("express");
const bodyParser = require("body-parser");

// Initialize a new Express application
const app = express();

// Configure the application to use Body-parser's JSON and urlencoded middleware
app.use(bodyParser.urlencoded({ extended: false }));

// Start the server
const port = process.env.PORT || 3002;

app.listen(port, () => {
console.log(`Server is running on port ${port}`);

Feature developmentโ€‹

Now we can use Aider to help us with the development of the app. Aider is a development accelerator that can help with code modifications and features development.

Aider interface:

โžœ aider
Aider v0.27.0
Model: gpt-4-1106-preview using udiff edit format
Git repo: .git with 4 files
Repo-map: using 1024 tokens
Use /help to see in-chat commands, run with --help to see cmd line args

Now we can generate feature for adding a new TODO item:


We can keep iterating by adding new features and testing. For example:


Code Iterationโ€‹

Finally, we can use GitHub Copilot to help us with the code iteration. GitHub Copilot is an autocompletion aid that can provide suggestions.

For example, here I want to log the GET request to the console, so I start typing:


And get autocomplete suggestions:


Obviously, the GAG stack is not the only set of tools, and the ones I've chosen might or might not have something to do with the resulting acronym. There is Devin, an open-source equivalent, Devina, that claims to be the first AI software engineer. There is Codeium, a free Copilot alternative. There are many other tools in this category, and the landscape is evolving rapidly.

Keen readers might have noticed that the underlying models used are OpenAI's GPT-3 and GPT-4. However, this is not a requirement. The tools can work with both local and remote models, paid and free. The choice of the model is up to the user.

So, are developers going to be replaced by AI? Are doomers or accelerationists right?


I think the answer is more nuanced. AI tools are here to stay, and they will be used to enhance our capabilities. The GAG stack is just one example of how AI can be utilized to assist us with software development.

As long as software development relies on human communication and creative collaboration, we will be talking about augmenting software development with AI rather than replacing it.