DEV Community: Vidip Ghosh

Building an AI SRE That Doesn't Just Detect Incidents - It Fixes Them

Vidip Ghosh — Sat, 25 Jul 2026 02:45:48 +0000

What is AegisSRE?

It is an enterpise-grade, governed autonomous systems that combines observability, AI reasoning, and governed automation to resolve production incidents with minimal human intervention.

Imagine it's 2 AM & our production service suddenly goes down. Traditional observability tools help engineers identify what went wrong through console logs, metrics and traces, but engineers still spend valuable time investigating the root cause, selecting the right runbook, executing remediation, and verifying recovery.

AegisSRE goes beyond observability. It detects production issues, raises incidents, diagnoses the root cause, selects the appropriate remediation, requests human approval for risky operations, executes the runbook, and independently verifies that the system has recovered.

Architecture

AegisSRE follows a multi-agent orchestration (Agent-to-Agent) architecture, where each specialized agent performs a single responsibility within the incident lifecycle. Each agent has a well-defined role:

Coordinator: Orchaestrates the complete incident workflow.
Classifier: Categorizes the incident and identifies the affected service.
Diagnosis Agent: Analyzes traces, logs and telemetry to determine the root cause.
Planning Agent: Selects the most appropriate remediation runbook.
Approval Agent: Enforces governance by requiring human approval for medium and high-risk actions.
Execution Agent: Executes the selected runbook in a controlled environment.
Verification Agent: Confirms that the remediation successfully resolved the incident before closing it.

Use of Signoz

SigNoz serves as the observability backbone of AegisSRE. Every workflow stage is instrumented using OpenTelemetry, allowing SigNoz to capture and visualize:

Distributed traces
Workflow spans
Application logs
Execution timings
Runbook execution details

How Aegis Uses SigNoz?

Aegis uses Signoz as the observability backbone of the autonomous incident response workflow.

Every workflow stage is instrumented using OpenTelemetry, allowing SigNoz to collect and visualize the complete lifecycle of an incident.

During an incident, Signoz provides:

Distrubuted traces to allow end-to-end execution of the workflow.
OpenTelemetry Spans for every AI agent, including coordinator, diagnosis, planning, approval, execution, and verification.
Application logs generated during diagnosis, remediation, and verification.

MergeGuard

Vidip Ghosh — Thu, 29 Jan 2026 17:22:06 +0000

This is a submission for the GitHub Copilot CLI Challenge

What I Built

MergeGuard is a proactive CLI tool designed to eliminate "Conflicts."

Problem: While working in a team we always needed to verbally communicate regarding the changes & ask other developers to take a pull OR while taking pull, unaware of changes we end up getting conflicts. So, here MergeGuard comes into play.

Key Features:
Real-time Watch Mode: Continuously monitors your active branch for remote updates.

Line-Level Analysis: It identifies the exact line-range overlaps between your local work and the remote.

AI Risk Assessment: Uses Llama AI to analyze the codebase and classify the conflict risk as NONE, LOW, MEDIUM, or HIGH.

Resolution Strategies: Provides AI-generated suggestions on how to approach the merge before it even happens.

Demo

Demo Video: Click Here

My Experience with GitHub Copilot CLI

MergeGuard integrates GitHub CLI (gh) to retrieve commit metadata (author + commit message) and combines it with git diffs to proactively detect merge conflict risk.

The Evolution of Sequential Learning Models: RNN LSTM Transformers

Vidip Ghosh — Tue, 11 Nov 2025 15:05:27 +0000

Recurrent Neural Networks (RNNs) are used for processing sequential or time-series data by maintaining a hidden states that captures memory of past inputs within a sequence.

They have a feedback loop, meaning each output depends on the previous inputs — making them ideal for tasks like sentiment analysis or speech recognition.

However, RNNs face key challenges:

Vanishing gradient problem – they struggle to learn long-term dependencies.
Limited memory – as sequences get longer, they gradually forget earlier information.

Long Short-Term Memory (LSTM) networks come into the picture as an improved version of RNNs, designed to solve the problems of vanishing gradients and long-term dependency.

Unlike traditional RNNs, LSTMs have a memory cell that helps retain important information for longer durations — allowing them to "remember" context across longer sequences.

Architecture Overview:

Forget Gate: Decides which part of the previous cell state to keep or forget.
Input Gate: It is actually the one that adds new information into the memory cell.
Output Gate: Controls what information is passed to the next step.

Real use case: Grammar correction tools like QuillBot, which rely on understanding long text dependencies to rephrase sentences accurately.

However LSTM face key challenges:

Computationally Expensive
Requires more memory.

LSTMs can’t analyze an entire sentence at once since they process information sequentially, step by step.
This is where Transformers come in — self-attention-based models designed for processing natural language efficiently. Unlike RNNs or LSTMs, Transformers process entire sentences in parallel, allowing them to capture context and relationships between words more effectively.
Examples: ChatGPT, BERT, GPT-4

Key Features:

Capable of performing Sequence-to-Sequence tasks (e.g., language translation).
Built using multiple encoder–decoder layers for deeper understanding.

🚀 From Algorithms to Neural Networks: ML vs DL Explained

Vidip Ghosh — Wed, 24 Sep 2025 16:12:34 +0000

We often hear Machine Learning (ML) and Deep Learning (DL) used interchangeably, but they aren’t the same.

🔹 Machine Learning (ML)

Machine learning learns from training data and then performs on new data.
It works well on structured data but classical ML models don't have layers, hence cannot work on complex data like image or do complex calculations.

🔹 Deep Learning (DL)

Deep learning is a subset of machine learning that uses many multilayered neural networks to model complex data. It uses artificial neural network.
Example: image classification, speech recognition, and natural language processing.

🧠 Core Neural Network Architectures

Artificial neural networks (ANN): ANN's work on the logic on how brain can perform calculations. It consists of 3 layers:

Input layer: receives data.
Hidden layer: process and learn patterns.
Output layer: generates results.

Convolution neural networks (CNN): CNNs are primarily used for image classification tasks.

Input layer: The input image (represented as a matrix of pixel values) is fed into the network.
Convolution layer: Here, most of the tasks takes place like extracting features.
Pooling layer:Reduces dimensions while keeping key features.
Activation layer: Activation functions (like ReLU) are applied after convolutional and fully connected layers to add non linearity to the model so that it can understand complex patterns.
Fully connected layer: Combines extracted feature for classification.
Output layer: Produces the final prediction (e.g., softmax for multi-class classification, sigmoid for binary classification, linear for regression).

Recurrent neural networks (RNN): It is used in sequential data. It requires memory, remembering past data etc. It is used in tasks like Natural language processing, Stock price prediction etc.

RNN is used for tasks like predicting next word.

💡 AI in Healthcare: Beyond Productivity Tools

Vidip Ghosh — Sun, 14 Sep 2025 17:18:01 +0000

We’ve all used LLMs like ChatGPT, Llama, and Generative AI for day-to-day tasks—boosting productivity, creativity, and even generating new ideas. These models can summarize long documents, analyze PDFs, interpret links, and generate text with ease.

Then we talked about fine tuning AI models according to our requirement. But AI isn’t just limited to productivity—it’s also transforming medicine.

For example: In healthcare, analyzing CT scans and medical reports is time-consuming. AI-powered agents trained on biomedical data can assist doctors by summarizing scans, flagging anomalies, and suggesting possible diagnoses—helping them save time and focus more on patient care.

Some specialized AI models in this field include:

ContactDoctor/Bio-Medical-MultiModal-Llama-3-8B-V1
Med-PaLM 2 (by Google DeepMind)
Llama3-OpenBioLLM-8B
TaozhiyuAI/OpenBioLLM-Llama-3

From Assistants to Decision-Makers: AI Agents vs. Agentic AI

Vidip Ghosh — Sat, 13 Sep 2025 19:49:28 +0000

When it comes in terms of automating stuff and decision making, AI agents & Agentic AI are often mentioned together- but they don't work the same way.

AI Agents

AI agents are task oriented i.e. it can only be used for specific task.
Designed to perform tasks independently.
Cannot help in decision making.
Ex: Email organizer → classifies emails into categories (spam, promotions, important).

Agentic AI

These are multi specialized agents that work together simultaneously. Capable of performing multiple tasks.
Can communicate, collaborate, and adapt to changing situations i.e. agentic AI are capable of learning from user interactions.
Ex: Healthcare agent which can both analyze documents as well as suggest us treatments.

👉 In short:
AI Agents = Specialists (narrow tasks, limited adaptability)
Agentic AI = Collaborators (multi-tasking, adaptive, decision-making)

Bias, Variance, and the Art of Building Better ML Models🎯

Vidip Ghosh — Fri, 12 Sep 2025 19:49:08 +0000

Machine Learning is about training models on a dataset, letting them learn patterns, and then testing them on unseen data.

Two important parameters to consider are:

🔹 Bias & Variance

High Bias → The model is too simple. It misses important patterns and performs poorly on both training and testing data.

High Variance → The model is too complex. It performs very well on training data but fails to generalize to unseen testing data.

⚡ Overfitting → High variance + low bias
Ways to reduce: choose only relevant features, reduce model complexity, or use techniques like regularization.

⚡ Underfitting → High bias + low variance
Ways to reduce: increase model complexity and use more training data.

🤔 Which is the best model?
The ideal model is one with low bias (captures patterns well) and low variance (generalizes well to new data).

From Generic to Specific: Making AI Work for Your Domain

Vidip Ghosh — Thu, 11 Sep 2025 19:10:37 +0000

Generative AI is trained on vast amounts of data and is widely used for:
✅ Generating new ideas
✅ Writing & debugging code
✅ Automating repeated tasks
✅ Image generation
…and much more.

But what if we want AI to answer questions on a specific topic?

Education → AI agents fine-tuned on a university’s curriculum to help students with course-specific Q&A.
Travel → AI trained on local attractions, food options, and transport to suggest personalized itineraries.

This is where two powerful techniques come in:

🔹 Fine-Tuning – retraining a Large Language Model (LLM) so it learns domain-specific knowledge and responds only in that context.

🔹 Retrieval Augmented Generation (RAG) – connecting an LLM to external datasets/documents so it can “retrieve” relevant, up-to-date information before generating responses, leading to more accurate, context-specific and reliable outputs.

Together, these approaches make AI more focused, reliable, and domain-specific—transforming it from a general assistant into a specialized expert.

Exploring Generative AI: From Curiosity to Creation

Vidip Ghosh — Wed, 10 Sep 2025 18:12:32 +0000

Artificial Intelligence is all about training machines to mimic human like behaviour. We feed our data to machine & it performs a specific task.
For ex: To detect if an email is spam or not spam, we first train our models using historic dataset using machine learning algorithms like Logistic regression and then we test on new & unseen data.
Also, for image classification we use a technique called Deep learning. It can analyze the complex patterns. Convolutional Neural Networks (CNNs) are the key technique for image tasks.

Traditional AI systems have one drawback. Traditional AI is usually limited to prediction or classification tasks—it can’t create brand-new outputs beyond what it’s trained on. Here comes the generative AI, Large language models (LLMs) and ChatGPT.

Generative AI learns from the massive amount of datasets and then generate new contents like image, text, code, etc.

Generative AI capabilities:

generating new ideas
enhancing productivity as it speeds up debugging, gives suggestions as seen in GitHub copilot
automating repeated tasks

Using .env in React + Vite

Vidip Ghosh — Sat, 29 Jul 2023 09:41:32 +0000

I was using .env in my react app created using vite. I have a token which needs to be hidden.

As soon as I wrote process.env.AUTH_KEY, I got the following error:

Then on searching I came across the right way of using .env variables in React + Vite project.

**
.jsx

const READ_KEY = import.meta.env.VITE_GH_AUTH_KEY;
const octokit = new Octokit({ auth: READ_KEY });

**
.env
VITE_GH_AUTH_KEY = 'your-auth-token'
**

For reference, please refer the article: https://stackoverflow.com/questions/72628089/error-in-vite-preview-uncaught-in-promise-referenceerror-process-is-not-def

Solving Adding another git repository inside another repository

Vidip Ghosh — Tue, 13 Jun 2023 14:28:04 +0000

I recently came across the situation where we push the project and as soon as we run git add . command, we get the following warning:

And when we push this project, we find that a submodule has been pushed into the project repository. The outer git repository ignores the inner git repository and the inner git repository is called the module.

I came across the solution for solving this:
Step1: run command ls -a (In git bash terminal)
Step2: run command rm -rf .git (remove .git)
Step3: Again, run git add . and commit the changes.

https://gist.github.com/claraj/e5563befe6c2fb108ad0efb6de47f265

Hacktoberfest 2022 badge

Vidip Ghosh — Thu, 03 Nov 2022 07:05:04 +0000

I have been awarded with badge for successful completion of Hacktoberfest 2022 challenge.