This repository tries to solve docker compose deployment problems through an AI Agent.

Pre Requisites

Installation of docker

Install docker from here

Installation of docker-compose

Install docker compose from here

Now install requirements.txt

Which ever python installation you are running be it virtual environment or anaconda install pip install -r requirements.txt

Construct RAG

For generating RAG you need to clone some existing docker repositories

mkdir -p repos
cd repos
git clone https://github.com/Haxxnet/Compose-Examples.git

git clone https://github.com/docker/awesome-compose.git 

git clone https://github.com/ruanbekker/awesome-docker-compose.git

Select either offline or online model. It could be ollama or online using

openrouter free key.

Note

If you are using github codespaces then it is very easy to try this just use

chmod +x setup.sh
./setup.sh

Model selection

Offline models

Installation of ollama

install ollama from here

Install offline models of deep seek

In one terminal keep running the following

ollama serve 

Also download these models

ollama pull deepseek-r1:1.5b 
ollama pull deepseek-r1:7b
ollama pull deepseek-r1:32b
ollama pull deepcoder:14b 
ollama pull llama3.2:3b 
ollama pull llama3.2:1b
ollama pull deepcoder:1.5b
ollama pull gemma3:4b 
ollama pull gemma3:12b 
ollama pull gemma3:4b

After downloading test the models using by first downloading curl and then use

curl http://localhost:11434/api/generate -d '{
  "model": "deepseek-r1:1.5b",
  "prompt": "one word about me just one word",
  "stream": false
}'

Make sure it works and it keeps on running at 11434

Online model (free tier)

1. Go to openrouter
2. Create a free key for deep seek. You need to choose option DeepSeek-R1 (free)
3. Copy that key and paste in code snippet given below
4. Run the following code to make sure it works.

from openai import OpenAI

client = OpenAI(
  base_url="https://openrouter.ai/api/v1",
  api_key="<your_created_key>",
)

completion = client.chat.completions.create(
 
  model="deepseek/deepseek-r1:free",
  messages=[
    {
      "role": "user",
      "content": "What is the meaning of life?"
    }
  ]
)
print(completion.choices[0].message.content)

Once you make sure that works run the following in linux use

export DEEPSEEK_FREE_KEY=<your_key>

in windows use

set DEEPSEEK_FREE_KEY="<your_key>"

Executing the application

Use

python start.py

It will redirect you to http://localhost:8501

Features

There are various features currently in this agent.

Execution Enviornment

We can run the generated files in the chat interface itself.

Model selection

We can select different models.

Chat interface

We can interact with AI to deploy our solutions in the chat interface itself.

Deployment Environment creation

As soon as the response is created all the genrated files are saved in file system in the provided directory.

Feedback

If a deployment is unsuccessful we can get feeback from the process to forward to the AI to look at it.

Monitoring

If a service is running we can continously monitor it and inform the AI about any failure.

Future plans

Another AI agent for manipulating this services.

1. Generating curl commands for interacting with various services.
2. Getting metrics, logs from the services, resource usage and making a sense of the data fetched and also closely monitoring these services.
3. Devising strategies for recovery in case of failure.
4. Increase load on services for stress testing.
5. Testing of backup restore of databses if databases are involved.
6. ~~We are parsing the files generated by the model and saving them. The model ~~ could be made to generate the code for environement creation as well. Done

Saving working configuration in database.

1. ~~Saving the configurations generated in database ~~ Done
2. ~~implementing RAG ~~ Done

Moving to kubernetes after this.

Repeating all of this kubenetes.

Reinforcement learning

The feeback information can be used to update the model parameters. But more realistc one would be first storing the monitoring information to RAG and then thinking of reinforcement learning.

Implement a MCP server

~~ Done ~~

Contributions

Contributions are welcome. You can reach out to me on mail (awsrohanyadav@gmail.com) or my linkedin or open an issue or pull request.

Credits

A lot of chat interface code has been borrowed from here

Thanks @krishnaik Buddy.

And @dcrey7

Thank you buddies.