What If Your Coding Agent Could Also Be Your SRE?

Shivesh Raj — Wed, 07 Oct 2026 06:57:07 +0000

A few months back, I was setting up a Central Monitoring system for all my projects where I could do log monitoring, cost monitoring, compute utilisation by each project, and so on. For that, I had already set up an ML-based algorithm to detect anomalies in cost and RAM, which then sent me alerts through my Telegram bot. You can read about it in my blog.

But once I set my project in production, I was receiving tons of logs, whether they were Django-specific, Docker-specific, Nginx-specific, or Database-specific. I sorted them in my Grafana Dashboard, but again, the problem was the amount of logs for each project, and not all logs necessarily needed my attention. However, error ones or critical ones did need attention for continuous rectification of the project.

I was having two ideas to resolve this issue. The first was either to go with a Small LLM and build a RAG application for retrieval of related logs and summarise them for me on a daily basis. The second was to build an MCP server on this monitoring server at a different port, where I could have different projects listed as options, different technology stacks listed as options, and then connect that to any of my coding agents, where I could ask it whatever information I wanted, whether it be a daily report or a critical error.

The first solution seems fine, but the roadblock comes in front of me while training a RAG application: model selection, weight training, compute resources, etc. But still, it would be only a half addition to my project. It’s like a doctor giving me only the diagnosis after analysing the report without knowing my medical history.

So basically, it was missing the context of the code on which that application was built and the solution to rectify the error that arises.

That’s where I came up with the solution where I can build the MCP server with various options of projects and the tech stack used in them. Using the skills of that project specifically, I can make sure that in the .md, every time I open up the project, it will give me the report at certain intervals.

As that coding agent will be having the code as context on which the application or project is built, the same coding agent can help me to rectify the error. And as the MCP server will be there, basically, you have created a controlled, monitored, self-healing system.

Here, my problem of separate project views and summaries based on each tech stack, and only needed attention being given to any project, saved me a lot of things.

This project was particularly interesting because I was new to SRE, as I had created just the Dashboard for my ease and kept on adding things to it.

Comment me for the project link; I will provide the GitHub link to my repo.

AI Is Not the Solution to Every Problem: When Rule-Based Systems Make More Sense

Shivesh Raj — Sun, 27 Sep 2026 12:44:19 +0000

Extensive Use of AI for Simple Task Costing Token and Even Then Not Producing Favorable Results

Today AI has been the buzzword every industry that wants to solve their task with the utilisation of AI, whether it be LLM application for RAG on technical document to ask for solutions, whether it be code, idea, calculation etc., any diffusion model for design, poster generation and all, using image processing model like YOLO to solve problems like defect detection.

But recent developments I noticed that there is too much extensive use of AI for simple task which a simple rule-based program can perform better. This is a concerning matter as the cost of running any simple AI task is far more than running a rule-based program.
A Practical Example: P&ID to Bill of Material

For example, last month I witnessed a software demo of creation of Bill of Material from P&ID using YOLO which is great as it is a repetitive task if customer ask for continuous revision and human error comes in picture with that counting each and every equipment present there makes this task extensive.

The solution is there are software to generate P&ID and BOQ but restriction is the template on which BOQ is required and the template on which P&ID is created.

Coming back to the YOLO based solution it uses the object detection technique where each P&ID is given in image form to the model and each equipment is detected as an object obviously each equipment used is trained on that model.

This seems a very good solution to give the count of it. But the limitation is the line tracing, properties and other technical difficulties comes in place due to a P&ID being a multilayer and additional of new equipment will need training of it again plus AI being the buzzword and if it comes in picture the cost of the solution sky rocket.

A Rule-Based Approach

Now this problem can be solved by simple use of logic and understanding of drawing and Python.

As any drawing is basically a graph and the layers are vertically stacked 2D sheet at different coordinate from one dimension and weight are looking at drawing from other axis so different stack of 2D graph gives you a combination of a 2D drawing as now the layers are separated then comes one layer in that you can consider each block used or equipment used is a node which is connected to each other by a pipe.

Now the basic understanding of graph analysis by Python using library ezdxf just change DWG to DXF or any other format to DXF as DWG is proprietary and closed as per AutoCAD.

Now that can simply give you layer by layer JSON of each equipment and as Python is handy you can use pyxl like library to directly insert these values in your required Excel template or to generate a PDF in required format.

This simple rule-based solution can save cost, perform better than any AI based solution with that it saves you hours of training and computation cost.

AI Should Be Used Where It Makes Sense

That's why I think everyone should read or at least decision maker should read the Kaggle course on AI ethics because that gives you clear idea where to use, how to evaluate and finally decision making for the use case.

AI is the buzzword but it necessarily should not be the solution of every problem statement.

The simple task solution can be great with a rule-based program works better than AI based solution as I shown earlier and this is also to give an idea to solution builder that before making solution for any use case please evaluate every aspect of it.

AI is a double-edged sword so keep that in mind the computation cost, training time, data collection and so on.

The Role of LLMs in Rule-Based Solutions

As one more thing once you devised the rule-based solution and created the logic you can get that coded from any LLM that will be fast, time saving and useful addition in the process.

DEV Community: Shivesh Raj

What If Your Coding Agent Could Also Be Your SRE?

AI Is Not the Solution to Every Problem: When Rule-Based Systems Make More Sense