Forem: Krishnan R

AWS Cloud Practitioner Essentials Certificate Achieved!

Krishnan R — Tue, 12 Aug 2025 12:25:57 +0000

🚀 Achieved the AWS Cloud Practitioner Essentials Certificate!

I successfully completed the AWS Cloud Practitioner Essentials certificate from AWS! This achievement was made possible by an excellent free course that helped me solidify my understanding of AWS Cloud concepts, including core services, security, pricing, and architecture.

Computer Vision Algorithms Led AI — Until Transformers Took Over

Krishnan R — Thu, 31 Jul 2025 08:10:32 +0000

Computer Vision Algorithms Led AI — Until Transformers Took Over

Until 2017, most AI advancements were driven by breakthroughs in computer vision, largely powered by Convolutional Neural Networks (CNNs). Models like ResNet, YOLO, and Faster R-CNN enabled significant progress in tasks such as image classification, object detection, and segmentation.

The Turning Point: Transformers in 2017

In 2017, the introduction of the Transformer architecture through the paper "Attention is All You Need" marked a major shift in the AI landscape.

Originally designed for Natural Language Processing (NLP)
Led to models like:
- BERT (Bidirectional Encoder Representations from Transformers)
- GPT (Generative Pretrained Transformer)
- T5 (Text-To-Text Transfer Transformer)

These models achieved state-of-the-art performance in many NLP benchmarks and brought language models to the center of AI research.

Transformers Expand Beyond Text

Over time, the impact of Transformers extended beyond NLP:

Computer Vision:
- ViT (Vision Transformer)
- SAM (Segment Anything)
Multi-modal Models:
- CLIP (connects text and images)
- DALL·E (text-to-image generation)

These models demonstrate the flexibility and scalability of the Transformer architecture across vision, language, and beyond.

A Paradigm Shift in AI

The shift from CNN-dominated pipelines to Transformer-based architectures represents one of the most significant transitions in the history of AI.

What do you think?

Let me know your thoughts in the comments below.

AI #DeepLearning #Transformers #NLP #ComputerVision #BERT #GPT #ViT #CLIP #TechTrends

Instance Segmentation with Mask R-CNN (ResNet-50 + FPN) using Detectron2

Krishnan R — Sun, 20 Jul 2025 18:53:30 +0000

# 🖼️ Instance Segmentation with Mask R-CNN (ResNet-50 + FPN) using Detectron2

Today, I successfully ran an instance segmentation model using Mask R-CNN with the ResNet-50 backbone and Feature Pyramid Network (FPN), based on the config file:

mask_rcnn_R_50_FPN_3x.yaml.

🔍 Model Architecture Overview

ResNet-50: Backbone network to extract rich feature representations from the image.
FPN (Feature Pyramid Network): Improves feature maps at multiple scales for better detection of small and large objects.
Mask R-CNN: Builds on top of Faster R-CNN by adding a segmentation branch to predict masks at the pixel level.

✅ Key Learnings & Workflow

Understood how to use and modify model config files in Detectron2.
Explored the model loading process from pretrained checkpoints.
Ran inference successfully on a sample input and verified the output.

Learning MLOps by Building a Real-World Salary Prediction Pipeline (MLflow + FastAPI + Docker)

Krishnan R — Mon, 14 Jul 2025 16:58:08 +0000

🚀 Hi , I am Going to Build a Real-World Salary Prediction MLOps Pipeline with MLflow

Learn experiment tracking, model versioning, and deployment by building an ML pipeline that predicts employee salaries.

This project is the foundation of my 15-day MLOps learning sprint — and I’m building it in public on LinkedIn & GitHub!

👋 Why I Built This

To Learn Mlops Pratically So I decided to learn by building a real, end-to-end project that involves:

✅ Tracking models & metrics with MLflow
✅ Serving predictions via FastAPI
✅ Adding a UI using Streamlit
✅ Preparing for production with Docker & CI/CD
✅ Building in public to grow with the community

I chose a practical use case — predicting employee salaries — and turned it into a full MLOps pipeline that simulates real-world AI workflows.

This project kicks off my 15-day hands-on MLOps learning challenge, where I’ll:

Build and track experiments with MLflow
Serve the model via an API
Add a clean UI for users
Package and deploy the app to the cloud
Learn CI/CD automation for model updates

💼 LinkedIn

If you're learning MLOps too — join me! Let’s grow together.