Session 08 – Agentic AI DevOps: ML Q&A

Agentic AI DevOps: ML Q&A

Recap of the previous sessions

• Regression predicts a number (a salary). Classification predicts a category (loan approved or rejected).
• Every training project follows the same flow: analyse the data, engineer the features, split into X and y, train (fit), predict, and measure the score.
• The model is judged on test data, never on the data it was trained on.
• Real data needs cleaning first: spaces, duplicates, missing values, encoding of categories, and scaling of big numbers.

In this session

This was an open question-and-answer session. Instead of a new topic, it covered:

• The assignment: predicting the time to resolve an incident
• Building a GitHub portfolio
• How to use AI coding assistants while you are learning
• A FinOps agent for AWS and Kubernetes cost savings
• What comes next in the course

In one sentence: use AI assistants to get the code, but make sure you understand the problem and the concepts yourself.

1. The assignment: predict the time to resolve an incident

The assignment applies what has been covered so far to a DevOps use case.

The business problem. When a business application goes down, the incident is reported to the service desk or the operations team. They diagnose it, and if they cannot fix it, they escalate to L1, L2 and L3 teams. The first question is always: how long will it take to fix? Engineers give an estimated time (ETA), and the business wants that number too.

The ML model. Train a model on previously reported incidents. When a new incident arrives, the model predicts the time to resolve. For example, it might tell the team that this type of incident usually takes about 15 minutes.

The assignment has a dataset link, a description of the goal, and a list of high-level tasks. It is a simple model, built only from what has been taught so far. A fully real-time system would need tools that come later in the course.

What to submit. Write your Python code or Jupyter notebook and commit it to your personal GitHub account. Share the link, and add a comment on which model and parameters gave you the best accuracy, so that others can learn from it.

2. Build a GitHub portfolio

• Create a personal GitHub account and commit everything you practise.
• By the end of the course, aim for a couple of projects in your repository.
• Interviewers click the GitHub link. They look at how often you commit, what projects you tried, and how you learn. It shows your skills better than a résumé line.
• All the code from the sessions (datasets and notebooks) is in the course repository under V1. Clone it and practise with it.

3. Using AI coding assistants while you learn

Learners asked whether they can use an AI assistant to solve the assignment. If you gave an assistant the assignment link, it would return a complete solution in seconds, and it is usually quite accurate. But you are learning, not delivering to a business. If you receive a finished solution, you will rarely go back and digest the concepts.

Recommended approach

1. Understand the problem first. What are you solving, why, what benefit does it give, and what are the steps of the training flow? You must be clear on this yourself.
2. Ask for one step at a time. First: “help me explore this dataset”. Run the code and understand the result. Then: “help me remove duplicates”. Then the next step.
3. Ask only for the syntax and code, not for the concept. Nobody expects you to memorise syntax these days. Your managers also expect you to use agents to speed up coding.
4. Explore alternatives. Ask why the task is done this way, and whether a different algorithm could be used. This deepens your understanding.
5. Balance the speed. Writing everything yourself from the documentation teaches the most, but it takes time. Using an assistant for syntax only is a good compromise.

Good habits with any assistant

• Validate what it suggests. Do not trust a fix just because it was suggested.
• Assistants keep proposing improvements (“can I add this?”). Know what your project needs, or you will drift away from your goal.
• Discuss the solution first, and let the assistant write code only after you agree with the approach.
• Be as clear as you can in your prompt. A clear prompt gives better results.
• Free plans have usage limits, so you may run out of prompts quickly. Whether to pay for a subscription is a personal choice. Learners reported that different tools give different quality for the same task, so compare them on your own work.

4. A FinOps agent

A learner asked whether an agent can look across many AWS accounts and find ways to save money, for example by shutting down instances at certain times or moving to a smaller instance type. The answer is yes, and a demo will come later in the course. The idea has two levels:

Notes from the discussion:

• Tools such as Karpenter and the horizontal and vertical pod autoscalers help reduce cost, but you still have to choose sensible instance types and minimum capacity. Poor choices still waste resources.
• AWS has its own recommendation tools, such as Compute Optimizer. They cover cloud-level resources. They cannot see the spikes inside your Kubernetes workloads, so an agent that reads those workloads adds value. If AWS’s built-in tools are enough for you, use them. If not, build your own agent.

5. What comes next

• The sessions continue with the machine learning topics, then move to the transformer architecture, the base of LLMs, followed by generative AI, LLMs and agentic AI.
• There will be occasional tech talks on recent updates and news in the market.
• New learners should watch the earlier recordings and read the session notes to catch up.

Summary

Key takeaways

• The assignment is a regression problem: predict the time to resolve an incident from past incident data.
• Commit your work to GitHub. It is your portfolio.
• Use AI assistants for code, not for concepts. Work step by step, understand each result, and validate every suggestion.
• A FinOps agent can recommend savings at both the AWS and the Kubernetes level.

Quick revision

Remember: you decide the approach, and the assistant writes the syntax.