Conducted a comparative study of text watermarking techniques on transformer-based models (OPT-350M, GPT-2 Medium), evaluating trade-offs in semantic fluency, coherence, and detectability.

Implemented and analyzed Red List / Green List watermarking (token-level constraints) vs. Cluster watermarking (distributional control) across BLEU, ROUGE, and Perplexity metrics.

Designed and ran experiments with multiple hyperparameter configurations (γ, δ), leveraging PyTorch, HuggingFace Transformers, Scikit-learn, NLTK, and SacreBLEU.

Findings: Red/Green list methods excelled in semantic fidelity, while Cluster watermarking preserved coherence with lower perplexity.

Future directions: scale experiments to larger models (GPT-J, LLaMA-7B), test adversarial robustness, and combine watermarking with reinforcement learning.

Built a smart recipe recommendation system using Retrieval-Augmented Generation (RAG) and semantic search to match user-input ingredients with 180K+ real-world recipes.

Implemented dense vector search with SentenceTransformers and ChromaDB, enabling real-time, context-aware recipe suggestions without relying on keyword matching.

Designed a fully CPU-compatible pipeline, optimizing for low-resource environments by avoiding model fine-tuning and using lightweight, pre-trained embeddings.

Preprocessed and normalized a large-scale recipe dataset (Food.com), handling data cleaning, formatting, and conversion into LangChain-compatible documents.

Deployed an interactive cooking assistant UI on Hugging Face Spaces using Gradio, supporting natural language ingredient input and seamless user interaction.

Fine-tuned Meta’s BART-Large-CNN model on a custom dataset of emails and conversational texts to generate high-quality, context-aware summaries, achieving a ROUGE-1 score of 57, ROUGE-2 score of 33, and ROUGE-L score of 48.

Employed Accelerate for mixed-precision training and multi-GPU support, optimizing model training on limited hardware resources. Designed efficient data preprocessing, batching, and augmentation pipelines to train on 15,000 text samples, ensuring diverse and generalized performance.

Integrated linear learning rate scheduling with warmup and AdamW optimizer to stabilize training and improve convergence.

Published and version-controlled the trained model on Hugging Face Hub using Gradio, making it publicly accessible for real-world text summarization tasks.

Developed a high-speed, lightweight dog breed classification model using PyTorch, trained on 120 unique dog breeds from the Kaggle Dog Breed Identification dataset.

Implemented transfer learning with top-performing vision models (ConvNext Tiny, ViT-B-16, MaxVit-Tiny), refining performance on a dataset of 10,000 images. ConvNext Tiny was selected for its balance between accuracy and efficiency, achieving 97% accuracy and a 95% F1-score.

Applied data augmentation techniques and a learning rate scheduler to enhance generalization and prevent overfitting, optimizing model performance.

Deployed the model on Hugging Face Spaces using Gradio, enabling real-time user testing and interaction.

Built a food classification model using PyTorch and the Vision Transformer (ViT) architecture, utilizing PyTorch’s torchvision library for transfer learning with pre-trained weights. The model was trained for 20 epochs, achieving an overall accuracy of 86.32%, with 20 food categories from the Food101 dataset.

To improve the model’s generalization and prevent overfitting, Applied data augmentation (flipping, rotating images) and used a learning rate scheduler to adjust the learning rate over time. Early stopping was also implemented to avoid overtraining the model.

Deployed the model using Gradio on Hugging Face Spaces, allowing live testing and interaction with the extended classification model, making it accessible for anyone to try.