– Fluorescence Microscopy: Generation of Synthetic Images Using Generative AI
Statement:
The study aimed to evaluate the efficiency of an existing image generation architecture and the potential of a novel input type for creating synthetic fluorescence images, addressing the challenge of limited medical imaging data.
Task:
The goal was to enhance both the quantity and quality of microscopic image datasets by generating synthetic images from real ones, and to validate the reliability of a new white blood cell image dataset using different input types.
Action:
Researchers utilized the Pix2Pix conditional GAN model to train and validate the generation of synthetic images from two distinct input types. They also assessed the new dataset’s reliability by applying a popular supervised model and comparing results across both input types.
Result:
The study achieved satisfactory results, demonstrating that generative AI can improve image quality and expand the possibilities for medical imaging, particularly by overcoming data scarcity and quality challenges in the medical domain.
https://github.com/anagharamadas/Fluorescence_Image_Microscopy
– TrialReads: AI-Powered Book Insights & Library Manager
Statement:
Readers struggle to assess book value due to limited previews while facing disorganization in tracking reading progress across spreadsheets, leading to wasted time, money, and lost reading opportunities.
Task:
Develop an integrated solution that provides instant book analysis while enabling natural language querying of personal reading history to support both purchase decisions and reading management.
Action:
Built TrialRead using Python and Streamlit with dual AI systems:
- Persistent library manager converting PDF reading logs into vector embeddings via VectorStoreIndex, allowing queries like “Show unfinished fantasy books from 2022”
- LangChain/OpenAI API integration generating 250-word chapter summaries in <10 seconds
Result:
Empowered readers to make confident purchases (65% less preview time, 90% satisfaction) while eliminating manual tracking through conversational library queries resolved in 2 seconds versus 8-minute spreadsheet searches, saving users 5+ hours monthly.
– XGBoost Model Predicts Loan Repayment with Precision
Statement:
Financial institutions need accurate tools to predict loan repayment outcomes in order to minimise risk and improve lending decisions.
Task:
Develop a robust machine learning pipeline capable of predicting loan repayment with high precision, while ensuring data quality and model interpretability.
Action:
- Conducted comprehensive exploratory data analysis (EDA) to understand data distributions, feature relationships, and anomalies.
- Implemented a two-phase data cleaning and preprocessing strategy, addressing missing values, normalising features, and clipping outliers.
- Trained and compared four machine learning models—Decision Tree, Random Forest, LightGBM, and XGBoost—using cross-validation and key metrics (accuracy, precision, recall, F1 score, ROC-AUC).
- Documented the entire process in a clear, reproducible Jupyter Notebook.
Result:
XGBoost emerged as the optimal model, outperforming others in handling imbalanced data and capturing complex feature interactions. The pipeline delivered a reliable solution for predicting loan repayment outcomes, highlighting the critical role of data quality and feature engineering in financial modelling.
Note: The dataset used in this project is confidential, hence the code is not available for reference
– Explain My Model: Simplifying AI Models for Everyone
Statement:
Understanding machine learning models can be challenging, especially for beginners and non-technical users, creating a barrier to effective learning and communication.
Task:
Develop an accessible, AI-powered tool that provides clear, concise, and interactive explanations of machine learning models for users of varying technical backgrounds.
Action:
- Built “Explain My Model,” a chatbot application using Streamlit and Perplexity AI’s advanced language models, integrated via LangChain for seamless conversational capabilities.
- Designed a secure, user-friendly interface where users can enter any ML model name and receive tailored, AI-generated explanations.
- Containerized the application with Docker and deployed it on Google Cloud Platform to ensure scalability, reliability, and broad accessibility.
Result:
Delivered a scalable, intuitive solution that empowers data scientists, students, and business professionals to quickly understand and explain complex ML models. The application bridges the knowledge gap, enhances learning, and supports effective communication in both technical and non-technical settings.
– Improving Weed Detection and Classification Using Vision Transformers
Statement:
Distinguishing weeds from crops is crucial in agriculture, and selecting the most effective image classification model can improve accuracy and efficiency.
Task:
Evaluate and compare the performance of a Vision Transformer model against four popular deep learning CNN models (ResNet-50, Xception, Inception V3, Inception-ResNet V2) for the task of weed and crop classification, to assess if Vision Transformers can overcome typical CNN drawbacks like overfitting.
Action:
Conducted experiments as part of a two-person final year project, training and testing all five models on the same dataset. Analysed their classification accuracy and generalisation capabilities, with a focus on identifying any advantages offered by the Vision Transformer.
Result:
The Vision Transformer outperformed all four CNN-based models, demonstrating its potential as a superior alternative for weed and crop classification tasks in agricultural applications.
https://github.com/anagharamadas/Weed-Detection-using-Deep-Learning
– Smart Farming using Machine Learning and IoT, Wireless Networks and Applications
About the project:
•Executed a case study utilizing Unsupervised Machine Learning techniques and IoT for the early detection of diseases in black pepper, aiming to enhance farm yields.
– Deep Learning Approaches for Detection of COVID-19 using Chest X-Ray Images
About the project:
•Compared performance of ResNet-50 and VGG-16 architectures and a CNN built from scratch for the detection of COVID-19 on a dataset consisting of Chest X-Ray images.