Hi. I'm Harun Kalkanci.

Efficient Retrieval-Based QA System with LangChain and Hugging Face

In this project, I share the results and implementations of building a powerful retrieval-based question-answering system using LangChain and Hugging Face. I walk you through the process step-by-step, leveraging the capabilities of state-of-the-art language models. For this project, I used markdown files from the Weights & Biases documentation as the data source.

• Loading and Preparing Data
• Splitting Documents into Sections
• Generating Embeddings
• Storing Embeddings in Chroma
• Alternative Solution: Prompt-Based Approach
• Creating a RetrievalQA Chain
• Running the Query

Technologies: LangChain, Hugging Face, Python

Practical Applications: Enhanced information retrieval, improved user experience, and efficient knowledge management.

Automated Image Captioning with AI

The Automated Image Captioning with AI project is designed to automatically generate descriptive captions for images extracted from a given website URL or a local folder. Leveraging state-of-the-art machine learning models, this project provides an interactive web interface using Gradio, making it easy for users to input a URL or select images from a folder, generate captions, and save the results in any format to enable modifying later as well.

Development Process:

• Situation: A news agency publishes hundreds of articles daily, each containing several images relevant to the story. Writing appropriate and descriptive captions for each image manually is a tedious task and might slow down the publication process. The agency needed a solution to expedite this process while ensuring the captions were accurate and contextually relevant.
• Task: Develop an automated image captioning tool that could generate descriptive captions for images extracted from a given website URL or a folder. The tool needed to be user-friendly, efficient, and capable of producing high-quality captions that enhance accessibility and improve SEO.
• Action: Implemented an automated image captioning program that works directly from a URL or a folder. The user provides the URL or selects images from a folder, and the code generates captions for the images found. The output is a text file that includes all the image URLs or image file names along with their respective captions.
• Result: By integrating this automated image captioning tool, the news agency is able to expedite its publication process significantly. The tool ensures that all images come with appropriate descriptions, enhancing accessibility for visually impaired readers and improving the website's SEO. This broadens the agency's reach and engagement with a more diverse audience base.

Technologies: Gradio, Hugging Face, Python

Practical Applications:

• Enhanced Accessibility: Helps visually impaired individuals understand visual content through descriptive captions.
• Improved SEO: Assists search engines in identifying the content of images, improving the article's SEO.
• Content Discovery: Enables efficient analysis and categorization of large image databases.
• Social Media and Advertising: Automates engaging description generation for visual content.
• Education and Research: Assists in understanding and interpreting visual materials.
• Data Organization: Helps manage and categorize large sets of visual data.
• Time-Saving: Automated captioning is more efficient than manual efforts.
• User Engagement: Detailed captions can make visual content more engaging and informative.

Transfer Learning for Scalable Deployment

This project demonstrates transfer learning using ResNet18 for image classification. The model was trained on a dataset of hymenoptera images and deployed using a Flask application.

Technologies: Python, PyTorch, ResNet18, Flask, Docker

Practical Applications: This project showcases the use of transfer learning to create a scalable image classification model. By leveraging a pre-trained ResNet18 model and fine-tuning it on a custom dataset, the project demonstrates how to achieve high accuracy in image classification tasks. The deployment of the model in a Flask application provides a user-friendly interface for image classification, making it accessible to a wide range of users.

AI Image Classification for Quality Control

Problem we are trying to solve: To build a Machine Learning model, that can evaluate images as good or bad, enabling the company to use the model for quality control purposes. This is a project to classify images from QA dataset from the AI4IM project.

Challenge => To build a single class classifier that has trained on enough good images so that it can detect bad images by only looking at them as seeing they are not good images, so that classify them as 'not good'='bad' images

Technologies: CNNs, Transfer Learning, Python

Practical Applications: Enhanced quality control, improved product consistency, and efficient defect detection.

Earthquake Prediction with AI

Used machine learning to predict earthquake magnitudes with an MAE of 0.39.

Technologies: Machine Learning, Python, Jupyter Notebook

Practical Applications: Enhanced earthquake prediction, improved disaster preparedness, and efficient risk mitigation.

Methods: Random Forest Regression, Gradient Boosting Regression, Support Vector Regression (SVR) (regression version of Support Vector Machine (SVM)), Neural Networks using TensorFlow/Keras, XGBoost Regression

Data: The project uses earthquake data from Turkey, including features like date, location, latitude, longitude, magnitude, depth, and more. The dataset is available in the data directory and in dataset location: Turkey 20 Years Earthquakes CSV. Please unzip the dataset and place it in the root directory.