Dr.Basanth Kalanoor
I am a data scientist with 3+ years of industrial experience in product development and management. I have a PhD in Physics from IIT Madras, where I managed a research project on quantum optics and conducted various experiments on photon interference. I have strong skills in data analysis, machine learning, programming, and communication.
View My LinkedIn Profile
Selected projects in machine learing, Computer Vision and NLP
Understanding The Disconnect between Skills and Jobs in Ghana usning NLP
Using web scrapping techniques, collected job description and candidate resume data from relevant websites. Fine-tuned DistilBERT, an NLP model, using the job description data to classify candidate resumes into different job categories based on their text. This project enabled fast and accurate analysis of large text data sets to extract insights from job description and candidate resume data. It also showcased my technical skills in NLP and my ability to use advanced tools and methods to solve complex data science challenges.
deployed steamlit application: https://nigeriajobskillgap.streamlit.app/
End-to-End AI/ML project for detection of chicken diseases using images of poultry fecal as the basis for its detection
In order to diagnose poultry diseases for small- to medium-scale poultry farmers, images of poultry fecal are used. The images are classified as "Coccidiosis", "Healthy", "New Castle Disease", and "Salmonella". The Python code is developed in a modular design based on the OOPS principle, and DVC is used for model and data monitoring, as well as versioning. The project has been containerized and deployed to AWS.
https://github.com/basanthsk/chicken-disease-classification
Deep Reinforcement Learning based Recommender System
In this project, we aim to implement a deep reinforcement learning based recommender system, inspired by the paper Deep Reinforcement Learning based Recommendation with Explicit User-Item Interactions Modeling by Liu et al. We use the DDPG algorithm, which combines actor-critic and deterministic policy gradient methods, to learn a policy that maximizes the expected reward for recommending items to users. To handle the high-dimensional and sparse user-item interaction data, we add a state representation module that transforms the data into a trainable state vector for the RL algorithm. This is not the official implementation of the paper, but our own attempt to reproduce and extend their work.
https://github.com/basanthsk/RL_DDPG
Improving Cognitive Abilities Through an AI-Driven Gamified Training System
The goal of the project from CogniFit was to create an AI model that can tailor a personalized and effective training plan for individuals to boost their cognitive skills. The AI model will determine the optimal next training session for each user, based on their cognitive profile and preferences. The AI system will suggest the best 2 games and their difficulty levels for the next session, with the goal of maximizing user engagement and cognitive enhancement. CogniFit will provide inputs such as user demographics, cognitive assessment results, and a pool of more than 50 games with varying difficulty levels. The system will improve with increased usage, function as an API, and be flexible enough to integrate new games into the learning model.
- Designed and implemented a reinforcement learning and collaborative filtering based recommendation system.
- Built an API to deploy the model and deliver the optimal next training session for each user.
- The AI system dynamically selected the best 2 games and their difficulty levels to maximize user engagement and cognitive improvement.
- Leveraged inputs such as user demographics, cognitive assessment results, and a pool of more than 50 games with varying difficulty levels.
- Developed a system that was adaptable to incorporate new games into the learning model.
Deploying an Accurate Classifier to Stop Online Violence Against Children using NLP
The objective of the team is to prevent online violence against children by using an effective classifier to detect grooming behavior in online chats. If the detected behavior reaches a certain threshold, the system will take action depending on the platform and intervention objectives, such as warning the child through a chatbot, alerting a moderator, or shutting down the chat altogether.
- Build on existing data and annotate additional sentences to reach a target of 100,000 annotated sentences with risk levels.
- Annotate new data made available by law enforcement.
- Scrape user data from online sources.
- Create a language model for grooming behavior detection and label it as non-risky, potentially risky, or risky.
- Test the data on various models and conduct ablation studies.
- Deploy the system as an API.
Predicting Short Term Traffic flow congestion on Urban Motorway Networks
The Smart-Traffic system is designed to predict the congestion levels of road traffic in real-time by utilizing traffic camera images and an EfficientNet image classification algorithm. The images are annotated and fed into the algorithm, which results in an accurate categorization of congestion levels as low, medium, or high. The system further enhances its predictive accuracy by comparing representative images and incorporating feedback from administrators. In addition to this, the project aims to expand its scope by obtaining API traffic image data for new cities and improving the model's accuracy by eliminating camera object noise.
AI for Sustainable Agri-Food Systems: Use of Satellite Imagery
The aim of this project is to develop an AI model that enhances the agricultural decision-making process by using satellite imagery to measure sustainability indicators. This will help address issues related to agricultural production, food consumption, and climate change sustainability, and support the adoption of sustainable practices by private and public stakeholders in the agri-food sector. By leveraging the power of AI and satellite imagery, this project aims to provide actionable insights and data-driven solutions to improve the sustainability of the agri-food value chain.
- Relevant data sources, including satellite imagery and other sustainability indicators, have been identified and gathered
- The data has been preprocessed and cleaned to ensure its quality and usability
- An AI model has been developed to estimate crop yield based on crop type and weather conditions
- The model has been tested and validated to ensure its accuracy and effectiveness
- The AI model has been integrated into the agricultural decision-making process to provide actionable insights and data-driven solutions.