Astha Soni

• Developed and fine-tuned MMBERT models for AI safety and classification tasks at Zscaler. • Created data generation pipelines utilizing multiple APIs to enhance data quality and accessibility. • Implemented comprehensive training and evaluation workflows in Python, ensuring robust model performance.

- Developed an audio guardrail system for LLM misuse prevention by fine-tuning Wav2Vec2 audio classification models with Hugging Face Transformers, achieving 85% jailbreak detection accuracy across diverse audio inputs. - Synthesized 5,000+ comprehensive audio samples from scratch using Google Cloud TTS, with manual labeling, and trained classification models on GPU-backed AWS EC2 instances, improving robustness by 18% compared to baseline. - Designed a robust and rigorous evaluation pipeline using Coqui XTTS voice cloning; validated performance across 4 benchmark datasets, achieving 86% F1-score and ensuring reliable cross-speaker generalization.

• Conducted comprehensive healthcare data analysis using Electronic Health Record (EHR) data to identify disease patterns. • Developed and evaluated machine learning models, including Random Forest and Decision Trees, to analyze clinical outcomes. • Enhanced model performance by addressing class imbalance issues through SMOTE oversampling and comparative evaluation metrics. • Utilized statistical methods such as univariate analysis and Cox proportional hazards modeling to assess risk factors and patient outcomes.

- Conducted data analytics with Market Basket Analysis and Cohort Analysis using python, leading to 13% increase in sales while utilizing Microsoft Power BI for visualizations of sales and Excel for intial data exploration. - Automated the data pipeline for reading raw data from an AWS S3 bucket, transformed it with business logic using Apache Spark, and saved the data in Databricks tables for streamlining data workflows for efficient analysis. - Built a rule-based chatbot, leveraging Google Dialogflow for NLP, PyCharm and FastAPI for backend processing, and MySQL for data management which led to a 20% increase in customer engagement and enhanced website traffic.

• Applied six machine learning models including Logistic Regression, SVM, KNN, Decision Tree, LightGBM, and XGBoost on PCA-transformed data from 28,000+ users for credit card fraud detection, working in collaborative environment. • Evaluated models on precision, recall, and F1-score, with LightGBM achieving 95% accuracy; performed XAI framework like SHAP, LIME and DALEX to enhance the interpretability and derive valuable insights for fraud prevention.

- Worked on financial analysis and rating recommendation of Tesla’s financial model, producing insights on 51.4% growth in 2022, and a 6.7% net profit margin, contributing to a winning presentation in a competitive setting. - Created a knowledge-based AI for Minesweeper, which strategically made decisions and drew inferences. Additionally, acquired sessions on Deep Learning, Computer Vision, machine learning and Lean & Agile principles.

- Developed the backend of a client-focused e-commerce website using Django, implementing robust features such as product catalog, user authentication, shopping cart, and secure payment processing, serving over 10,000 active users. - Leveraged Django’s Model View Template architecture and incorporated agile methodologies and utilized big data technology such as Hadoop, to enhance data processing capabilities, system performance and scalability.

Introduction to Probability Data Analysis Reinforcement Learning Programming Abstraction Statistical Learning Statistical Computing