In Winter of 2024, Cheenta Research School worked with 10 research groups from schools in India, US, Singapore and Middle East
Students from Grade 8 to 12 produced outstanding research papers.
Research Resources
Student Talks... exciting research
Research Seminar: AI for Social-Driven Crypto Pricing
Abstract: In this project, we try to understand how news headlines about cryptocurrency affect the price direction of Bitcoin. We collected real-time news using the CryptoPanic API and used different sentiment models like FinBERT and CryptoBERT to find whether each headline has a positive, negative, or neutral sentiment. Then we tested several decision models, such as Random Forest and others to see how well they can predict changes in Bitcoin pricing. We also used Google Apps Script to automate daily headline collection. By changing the sentiment models and the decision models one at a time, we aim to find the best combination for making accurate predictions about Bitcoin’s price movement.
AI for VVIPS Driven Asset Pricing
Abstract: This study investigates the impact of Elon Musk’s tweets on stock price fluctuations, leveraging artificial intelligence (AI) techniques to analyze sentiment and predict short-term market behavior. As a high-profile figure, Musk has demonstrated a unique ability to influence financial markets through social media, particularly platforms like Twitter (now X). Using natural language processing (NLP) and sentiment analysis, we examine a dataset of Musk’s tweets alongside corresponding stock price movements of companies closely associated with him, such as Tesla. Our findings reveal a significant correlation between tweet sentiment and immediate stock performance, with positive sentiment often linked to price increases and negative or ambiguous sentiment correlating with volatility or declines. However, the study also finds that sentiment alone does not fully explain market responses; similar messages from less influential individuals do not yield comparable effects. This suggests that the identity and perceived influence of the message source are critical factors in market reactions. These results highlight the limitations of sentiment-only models in financial forecasting and underscore the need for context-aware systems that incorporate social influence and credibility metrics. The study contributes to the growing intersection of behavioral finance, social media analysis, and AI-driven market prediction.
AI for Games (Basketball)
This paper discusses the methods and results of our analysis of the factors in determining the winner of the popular online game Basketball Heads. We collected data on 28 players and used decision trees to determine the most important factors. Our analysis shows that ball possession is the most important factor, with an ~80% accuracy in determining player outcomes.
AI for Taste Detection
In this project, we're using AI and computer vision to estimate fruit sweetness (measured by brix index) from simple photographs and multispectral images. We've developed machine learning models that analyze images of grapes to predict their sugar content, and we've gone further by testing whether including additional information like grape variety and harvest time can make these predictions even more accurate.
What makes this research special is that our models work with ordinary RGB images, meaning farmers could potentially use just a smartphone camera for quality checks. We put this to the test with five different grape varieties including Itum, Autumn Royal, and Crimson, examining how visual features like color and texture relate to actual sugar content. The most promising aspect is how this technology could help farmers quickly assess fruit quality right in the field, without needing expensive lab equipment or complicated processes.
What makes this research special is that our models work with ordinary RGB images, meaning farmers could potentially use just a smartphone camera for quality checks. We put this to the test with five different grape varieties including Itum, Autumn Royal, and Crimson, examining how visual features like color and texture relate to actual sugar content. The most promising aspect is how this technology could help farmers quickly assess fruit quality right in the field, without needing expensive lab equipment or complicated processes.
Homomorphic Broadcast Encryption
This paper presents a unified cryptographic framework called Ho-
momorphic Broadcast Encryption (HBE), which combines homomor-
phic encryption for privacy-preserving computation with broadcast encryption for selective access control. We implement and evaluate HBE in cloud and IoT environments, enabling secure data aggregation and controlled result distribution. Experimental results on a simulated 9-node cloud with 12 users demonstrate efficient encrypted computation, dynamic user revocation, and minimal latency. HBE proves effective for real-world applications such as healthcare analytics and financial risk assessment, offering a practical solution for secure collaborative processing in untrusted systems.
momorphic Broadcast Encryption (HBE), which combines homomor-
phic encryption for privacy-preserving computation with broadcast encryption for selective access control. We implement and evaluate HBE in cloud and IoT environments, enabling secure data aggregation and controlled result distribution. Experimental results on a simulated 9-node cloud with 12 users demonstrate efficient encrypted computation, dynamic user revocation, and minimal latency. HBE proves effective for real-world applications such as healthcare analytics and financial risk assessment, offering a practical solution for secure collaborative processing in untrusted systems.
AI for Ocean Trash Detection
Abstract - This research project aims to investigate and develop a highly efficient method in order to automate water quality assessments using readily available satellite imagery combined with a custom machine learning model. The study presents a deep learning approach for the seamless detection of water pollution from satellite imagery using a convolutional neural network (CNN) architecture integrated with a Class Activation Map (CAM) to enhance interpretability. The CNN architecture involves three convolutional blocks, global average pooling, and a fully connected classifier. Trained on three different datasets of polluted and natural water bodies, the model achieved high accuracy, demonstrating its usefulness in image classification tasks. Compared to existing solutions that contain millions of parameters, this model uses far fewer parameters and has fewer layers, making it a viable option for pollution detection tasks in the future.
Why Research?
Advanced students apply for research projects. It helps them learn more, improve their CV and gives them an edge in university applications.
Talk to Advisor