Projects
● Engineered a detailed hierarchical taxonomy categorizing stress into physical, psychological, psychosocial, and psychospiritual dimensions, enhancing data annotation precision and interpretability.
● Designed and implemented advanced data collection protocols, utilizing real-time counselor-patient dialogues to generate high-fidelity simulation datasets for model training.
● Executed fine-tuning of LLAMA 2 models using custom-generated datasets, optimizing hyperparameters and employing advanced training techniques to improve conversational AI performance.
● Developed and integrated bespoke evaluation metrics, including sentiment flow, active listening skills, and empathy scores, to rigorously assess model output quality and relevance.
Bridging AI Dimensions: Small Model Precision Meets Large Model Depth in Therapy
Designing Healthcare AI Tool specifically for Non-English Speaking (NES) patients
● Conducted a hospital visit to gather real-world application insights for refining the AI tool.
● Presented technical solutions and interactive design prototypes to stakeholders.
● Selected and integrated reliable translation APIs like OpenAI and Google Translator for English to Non-English (mainly Spanish) translations, ensuring high accuracy and contextual relevance.
● Analyzed sample datasets to understand the text structure in both English and Spanish, providing a detailed foundation for translation accuracy and term identification.
StressMix Project: Benchmark Dataset for Stress Type Prediction from Social Media
● Shaped a nuanced stress taxonomy, sculpting a dataset for pinpoint stress categorization.
● Optimized stress categorization through dataset label linkage, addressing critical gaps, and implemented healthcare efficiency enhancements with our
dataset.
Building a fine-tuned recommendation system
● Developed a recommendation engine employing content-Based and collaborative filtering systems to predict customer preferences for content, products, or services.
● Implemented a recommendation system using the Single Value Decomposition (SVD) algorithm, a matrix factorization technique, to break down a user-item matrix and generate accurate predictions.
● Utilized the Yelp dataset, a collection of user-generated reviews, ratings, and metadata, encompassing various industries such as restaurants, shopping, and nightlife.
● Loaded input data from JSON files, preprocessed data to extract relevant features, chose the SVD algorithm for recommendation modeling, trained the model using training data, evaluated its performance with validation data, and utilized the trained model to make personalized recommendations for target user-business pairs in the test dataset.
Getting in Tune: Instruction Tuning of ChatGPT3
● Evaluation of prompt styles (templates, direct prompts, few shots, persona prompts) for large language models (LLMs) in terms of grammaticality and typicality.
● Discrepancies observed in prompt performance: some prompts excelled in grammaticality but lacked typicality, and vice versa.
● Prompts using predefined templates resulted in responses with low typicality scores, potentially limiting diversity and contextuality.
● Direct prompts, involving straightforward questioning, yielded the lowest grammaticality scores, indicating potential limitations in generating syntactically accurate responses.
● LLMs demonstrated challenges in comprehending specific prompt categories, highlighting areas for improvement in their understanding and response generation capabilities.
Indian Sign Language Conversion using Machine Learning
● Developed a real time gesture recognition of Indian Sign Language.
● Handled the development of the ISL dataset consisting of about 1500 images for alphanumeric characters and another 1500
images for words used in day-to-day conversations.
● Research and development undertaking to imply different methods for the project execution. As a result of which a TensorFlow object detection model, using the Single Shot Detector (SSD) algorithm was employed using Python programming language.
● VGG 16 layer in the SSD Model was exercised for feature extraction and the convolutional filters were employed for
prediction in real time. The predicted output was further converted into audio by using the gTTS api.
Detection of Emotion from Texts Using Python
● Developed a tool to detect emotions in text taken from different tweets as the dataset having texts, numericals and emoticons. Tokenization was conducted after retrieving the dataset. The dataset thus created was applied to identify the emotion words from the entire text.
● Executed the analysis of the intensity of the emotion words which were followed by the negation check.
● Applied the VADER sentiment Analysis method which is both the lexicon and rule based to get the positive, negative and neutral emotions as results.
