Shri Harsha Adapala

The process involves understanding the semantics of the prompt, generating relevant lyrics, identifying the musical genre, and composing melodic and harmonic elements to create the final song.

◦ The methodology is divided into three main stages: Lyric Generation from Textual Prompts, Genre Classification of the Lyrics, and Chord Progression Generation Conditioned on Lyrics and Genre.

◦ Lyric Generation from Textual Prompts - Fine-tuned LLMs (FlanT5 and Llama models) generate lyrics from encoded prompts, with the LLM fine-tuned on a lyrics-prompt dataset for three epochs.

◦ Genre Classification of the Lyrics - Pre-trained models (BERT, DistilBERT, and RoBERTa) are fine-tuned for genre classification to classify the generated lyrics.

◦ Chord Progression Generation Conditioned on Lyrics and Genre - Combines a two-stage encoder-decoder model (DistilBERT and LSTM) with reinforcement learning (RLHF) to generate chord progressions based on lyrical context.

Exploring Multiagent Systems utilizing Reinforcement Learning within Cooperative and Mixed Cooperative and Competitive Environments.

◦ Solved Environments - Grid World Cooperative Environment, Mixed Cooperative and Competitive Environments Multiparticle Simple Adversary Environment.

◦ Implemented Algorithms - Decentralized Q-Learning, Correlated Q-Learning, REINFORCE, MultiAgent Deep Q-Learning, MultiAgent Deep Deterministic Policy Gradient.

Focus on developing a deep learning-based system for accurate identification and analysis of human actions in varied environments.

◦ Developed a sophisticated deep learning system tailored for accurate recognition of human actions across diverse settings, particularly for surveillance and sports analysis, utilizing a refined subset of 10 classes from the UCF101 dataset.

◦ Applied AlexNet3D and VGG3D models for advanced video data processing.

◦ Segmented the videos for an effective background and human masking

◦ Achieved notable metrics: AlexNet (up to 36% accuracy, 37.51% precision) and VGG (up to 39% accuracy, 35.26% precision) under varying conditions.

Explores the OpenAI Gym library, implementing the DQN algorithm as described in DeepMind's seminal paper, and subsequently improving the DQN algorithm for enhanced performance and stability.

◦ Implemented A2C

◦ Solved Environments - Cartpole, Lunar Lander, Inverted Pendulum, Bi-Pedal Walker

Explores the OpenAI Gym library, implementing the DQN algorithm as described in DeepMind's seminal paper, and subsequently improving the DQN algorithm for enhanced performance and stability.

◦ Implemented DQN and Double DQN

◦ Solved Environments - Cartpole, Lunar Lander

NovelConvo AI is an innovative Q&A and chat system tailored for classic book enthusiasts.

◦ Trained and Integrated Chatterbot, enhancing its capabilities with diverse data sets for natural and contextually relevant dialogues.

◦ Utilized DeBERTa-v3-large-mnli-fever-anli-ling-wanli to classify novels and prompts, streamlining user interactions.

◦ Developed an Inverted Index using Python and Flask, integrating Linked Lists for efficient data management.

◦ Implemented a Document-at-a-Time (DAAT) Boolean query processing system for advanced information retrieval.

◦ Integrated Open AI’s Retrieval Augmented Generation to the system to make responses more sophisticated.

◦ Developed a chatbot web application using ReactJs.

Language Translator from English to French.

◦ Constructed and deployed a Sequence-to-Sequence network using an encoding and decoding paradigm, utilizing LSTM nodes as the foundation. Implementation achieved a noteworthy test Bleu score of 0.365.

◦ Executed a bidirectional Sequence-to-Sequence network, incorporating both forward and backward encoding. Implementation, succeeded in a commendable test Bleu score of 0.369, highlighting the model’s proficiency in generating high-quality sequences.

◦ Performed a Sequence-to-Sequence network with bidirectional encoding and incorporated an Attention mechanism in the decoding process. This advanced architecture led to remarkable results, attaining a notable test Bleu score of 0.406. The model’s ability to generate accurate and coherent sequences was improved by the inclusion of Attention.

This project explores the efficacy of reinforcement learning algorithms in navigating complex grid environments with varying levels of determinism and stochasticity.

◦ Three distinct grid worlds are studied, ranging from deterministic to stochastic transitions and rewards.

◦ Implemented reinforcement learning algorithms include SARSA, Q-Learning, N-Step SARSA, Double Q-Learning, Deep Q-Learning, and Double Deep Q-Learning.

Implementation of an Inverted Index using Linked Lists and Flask in Python. It allows for efficient Boolean queries using a Document-at-a-time (DAAT) strategy, ideal for understanding Information Retrieval concepts.

◦ Developed an Inverted Index using Python and Flask, integrating Linked Lists for efficient data management.

◦ Implemented a Document-at-a-Time (DAAT) Boolean query processing system for advanced information retrieval.

◦ Crafted a text preprocessing module for document tokenization and normalization.

◦ Enabled interactive querying through a Flask web application, demonstrating web development and information retrieval skills.

Academic project on different image classification tasks.

◦ Developed and trained an initial AlexNet model, and then I modified the AlexNet architecture to enable the classification of three categories: Dogs, Food, and Vehicles. This modification resulted in impressive accuracy rates of 90% and 92.6% for the respective categories. Subsequently, I implemented the VGG-13 model, employing Mixed Precision Training techniques. This approach yielded a remarkable accuracy of 91.4%.

◦ Executed the implementation and training of a customized AlexNet model for the classification of Google Street View House Numbers, resulting in an impressive accuracy of 91.4%.

◦ Developed and trained a customized AlexNet model for classifying the OCTMNIST dataset, yielding an accuracy of 71%.

◦ Created and trained a customized AlexNet model to classify a 10-class ImageNet dataset, achieving an accuracy of 68.4%. Then, I implemented the VGG-13 model and employed Mixed Precision Training, which led to a notable accuracy improvement of 71.8%.

This project applies a Q-learning agent to develop a trading strategy that maximizes profit through stock trading.

◦ Leverage Q-learning to develop an optimized trading strategy for stock trading.

◦ Trained to learn stock price trends and make buy, sell, or hold decisions based on its learned strategy.

◦ Percentage return on investment (ROI) willl be Perfromance Metric.

Revamp football data management with a dynamic player database system.

◦ Created and maintained a Fifa23 Players database from scratch using database design theory. This involves data massaging using Python, modeling E/R diagrams, and translating it into a relational schema.

◦ Analyzed functional dependencies and normalized the tables to eliminate data redundancies and anomalies.

◦ Formulated and executed a variety of SQL queries to address common retail scenarios, and optimized queries for efficient data retrieval.

◦ Designed and created an interactive PowerBi Dashboard that provides visual insights into the database, showcasing different stats.