Student

Ashwini Ramesh Kumar

413-***-**** # ************@*****.*** ï linkedin.com/in/ashwini-ramesh-0509a71aa/ § github.com/a-sh28 Education

University of Massachusetts, Amherst (Expected) Aug 2023 - May 2025 Master of Science, Computer Science Amherst, MA

Coursework: Machine Learning, Advanced Natural Language Processing, Intelligent Visual Computing, Software Engineering, Reinforcement Learning, Distrubuted Computing Systems Sri Sivasubramaniya Nadar College of Engineering Jul 2019 - May 2023 Bachelor of Technology, Information Technology(First class with distinction) CGPA: 9.003/10 Head - Entrepreneurship Development Cell, Under Secretary General, Sponsorship - SSN MUN 23 Technical Skills

Languages: Python, Java, C, C++, HTML/CSS, JavaScript, SQL, PL/SQL. Technology/Tools: NLP, Computer Vision, Speech, Git, Docker, AWS, Data Structures, RaectJS, NOSQL, Express, Flutter, Firebase, Pandas, OpenCV, Seaborn, Unix, Linux, Jira, Tensorflow, Pytorch Experience

CitiBank Apr 2022 - Jul 2022

Summer 2022 Software Developer Intern Chennai, India

• Developed a lookup tool for Citibank’s corporate banking portal named ’CitiConnect for Files’, using ReactJS, Java-SpringBoot, and MySQL to visualize transactions, track payment statuses, and generate detailed financial reports.

• Achieved a 20% improvement in transaction visualization speed, resulting in faster decision-making for clients, and implemented a real-time payment status tracking feature, reducing manual effort involving 30 minutes of time on average per data pull.

• Utilized DevOps tools such as Git for version control, Docker to create reproducible development environments and Jenkins to streamline the development pipeline and enable Continuous Integration.

• Developed a Chatbot for banking applications, utilizing NLP techniques for pre-processing of queries and feed forward neural network to train the model. The user interface was built using Python - Flask. Verzeo Edutech Oct 2020 - Nov 2020

Machine Learning Intern Remote

• Implemented Ensemble Machine Learning Techniques, combining Support Vector Machines (SVM), Na ıve Bayes, and Logistic Regression algorithms for gender classification with a test accuracy of 98%

• Acquired proficiency in Data Analytics concepts, employing Python libraries including Pandas, Numpy, Scikit-learn, Pytorch and Seaborn facilitating exploratory data analysis. Projects

Fine-Tuning Intelligence: Exploring Dynamic Prompts in RLHF for Improved Alignment RLHF, LLMs

• Devised a custom RLHF pipeline to fine tune GPT2 -Medium, implementing dynamic prompts directed towards aligning the model on helpfulness.

• Employed SHP, HelpSteer, and RealToxicity datasets to develop a dynamic prompt generation pipeline. This involved pausing training to identify the model’s failure points and generating additional prompts based on these points to improve training in subsequent epochs.

• Performed LLM based evaluations on the baselines GPT2 Medium, Vanilla RLHF and GPT-2 SFT trained model using Gemini-Pro and other metrics such as Perplexity, Toxicity and generation lengths. Code Summarizer and Evaluator React JS, Flask, MySQL

• Built a code summarizer that can summarize, evaluate, and translate multiple code files.

• Utilised Google’s PaLM API through backend Flask Server to set up a pipeline to summarise, evaluate and return feedback and MyMemory API for translation.

• Conducted Unit Testing using React Jest, Integration Testing using Postman and Equivalence testing.

• Deployed the application on Google’s App Engine. Learning to play Cartpole and Lunar Lander Actor Critic, Proximal Policy Optimisation

• Developed a RL agent to learn the optimal policies in Cartpole and Lunar Lander MDPs using two policy gradient algorithms - Actor critic & Proximal Policy Optimization(PPO) in Tensorflow

• Conducted analytical study on the learning of the agents across the two algorithms, convergence to optimal policy, convergence speed, inferences from learning curves and PPO loss. Concluded on PPO to perform better for the above chosen MDPs.

Speaker recognition under Lombard and Normal conditions for 4 different emotions MLPs,Speech

• Curated a dataset with 20 speakers on 8 combinations of acoustical behaviour and elucidated strategies to filter voiced packets using Librosa

• Constructed 8 robust MLP models to recognise speakers under infused noise, using MFCC features, GridSearchCV with an average test accuracy of 96.23%

• Studied the impact of the infused noise and speaker emotions on the performance of the system using Audacity for detailed phonetic analysis and TensorBoard for visualizing model performance metrics. Gesture based Sterile Browsing System for Radiology Images Pixel Attention CNNs, Image Processing

• Devised strategies to filter the hand gesture from the input image and augmentation of the retrieved gesture to diversify the data.

• Built a Pixel Attention CNN model for classification of hand gestures and Python-Flask to build the interface.

• Received an excellent test accuracy of 99%, after three iterations of hyperparameter tuning. Achievements

• Placed 1st at district level, Interschool-innovation-challenge conducted by BiBox Labs.

• Won 3rd place at Inter School-College Coding Competition for programming in C conducted by Amrita College of Engineering, Bengaluru.

• Placed 2nd in Buildathon, Invente, an inter-college technical contest conducted at SSNCE, Chennai for proposing an optimized model for waste segregation.

• Cleared the technical round at Hack with Infy, a coding competition conducted by Infosys.

Contact this candidate