Nadeesha Vidyamali

My Contributions:

• Led data cleaning and preprocessing: handled missing values, standardized formats, and removed low-information columns to ensure data integrity.
• Engineered key features: computed patient age groups, length of stay, admission month/year, and billing categories to enrich analysis.
• Performed exploratory data analysis: visualized condition prevalence, billing trends, and appointment adherence across age, gender, and medical condition segments.
• Applied machine learning: one-hot encoded categorical variables, scaled numerical features, and trained a stratified random forest classifier to predict test results.
• Communicated findings: prepared comprehensive visualizations and a detailed report with recommendations for targeted preventive care and operational improvements.

Technologies Used:

My Contributions:

• Data Preparation & EDA: Cleaned and standardized 569 samples with 30 numeric features, dropped irrelevant columns, balanced classes, and visualized feature distributions and pairwise relationships.
• Feature Engineering: Applied StandardScaler, encoded labels, and performed stratified train–test splits with cross-validation to preserve class proportions.
• Modeling & Tuning: Implemented and tuned four classifiers—Logistic Regression, SVM, Decision Tree, and a two-layer ANN—with systematic grid searches (C, kernel, max_depth, layer sizes) and early stopping for the ANN.
• Reporting: Created comprehensive visualizations (feature distributions, performance bar charts, ROC curves) and delivered actionable recommendations for deploying the top-performing ANN in clinical settings.
• Evaluation & Analysis: Compared models using accuracy, precision, recall, F1-score, AUC, confusion matrices, and ROC curves; identified concave points, perimeter, radius, and area as key predictive features.

Technologies Used:

My Contributions:

• Developed an end-to-end deep-learning pipeline to automatically identify five key savannah species (buffalo, elephant, giraffe, zebra, rhino) from real-world camera-trap imagery, achieving robust performance across challenging lighting and environmental conditions.
• Preprocessing & Augmentation: Applied Gaussian noise reduction, resized/normalized all images to 128×128 pixels, and generated 5× augmented variants per image (rotations, shifts, zooms, brightness, flips).
• Trained the model, monitored performance, and applied techniques to prevent overfitting.
• Model Architecture & Training: Designed and trained a custom CNN in TensorFlow/Keras, implemented stratified train/validation splits, and monitored accuracy and loss curves over 10+ epochs.
• Evaluation & Visualization: Produced ROC and confusion-matrix plots to assess per-species performance, identifying avenues for further hyperparameter tuning.
• Reporting & Insights: Compiled a comprehensive report summarizing methodology, quantitative results (60% test accuracy), and recommendations for future improvements (additional data, advanced architectures).

Technologies Used:

My Contributions:

• Dataset Curation & Preprocessing: Collected 34k+ BSL images, applied MediaPipe key-point detection to crop and center hands, resized to 224×224, normalized pixel values, and used rotation/flip/brightness augmentations to improve robustness.
• Model Development & Fine-Tuning: Adapted a pre-trained EfficientNetB5 (ImageNet) by unfreezing its top 30 layers, added GAP → Dense(512, ReLU) → Dropout(0.5) → Softmax output, and trained with Adam (1e-4) over 10 epochs.
• Created a variety of plots (histograms, scatter plots, heatmaps) to represent data effectively.
• Evaluation & Analysis: Assessed performance via accuracy (91%), precision, recall, F1-scores, confusion matrices, multi-class ROC and precision-recall curves—identifying gestures (e.g. ‘O’, ‘N’) needing further data for disambiguation.
• Deployment & UX: Packaged the model into a Gradio web interface allowing drag-and-drop image uploads, real-time landmark visualization, and on-the-fly prediction—ensuring accessibility without local installs.
• Challenges & Future Work: Mitigated background noise and lighting variability through targeted augmentations; outlined edge-device deployment (Raspberry Pi) and live-camera support to broaden real-world applicability.

Technologies Used:

My Contributions:

• Data Preprocessing: Applied tokenization, lemmatization, cleanup (regex & stop-word filtering), and sentence segmentation to prepare raw text for downstream tasks.
• Text Summarization: Built and trained an LSTM‐based Seq2Seq encoder–decoder with attention in TensorFlow, achieving a ROUGE-1 score of 0.67 and BLEU of 0.22 for coherent story abstractions.
• Semantic Search: Generated sentence embeddings (all-MiniLM-L6-v2) stored in ChromaDB to enable fast, context-aware retrieval (most similarity scores >0.7) of relevant passages.
• Topic Modelling: Tuned a Gensim LDA model (5 topics) on TF-IDF–filtered text to uncover core themes—detective work, crime, dialogue—and visualized their distributions across the stories.
• Deployment: Packaged the pipeline into an interactive Anvil web app supporting on-the-fly summarization, search, and thematic exploration, with optimizations for inference speed (quantization, caching).

Technologies Used: