Vikas Thirumanyam

Btech Walleh in association with Teachnook | Remote

• Achieved 90%+ accuracy in sentiment classification using DistilBERT, deployed via Flask API.
• Reduced manual feedback analysis time by 40% through automated sentiment chatbot implementation.
• Improved customer sentiment understanding by 85% by providing instant analysis using the developed chatbot.
• Decreased API latency by 15% through optimized DistilBERT model deployment in Flask.

Ad Click Fraud Detection using Machine Learning

• Engineered an XGBoost model that identified click fraud with 92.5% F1-score, using SHAP for interpretable ML insights.
• Applied advanced feature engineering and SMOTE for class balancing across 100k+ records.
• Integrated real-time visual dashboards to monitor fraud trends, improving analyst decision speed by ~40%.

BERT-Powered LLM for Comment & Metadata Flagging

• Fine-tuned a BERT-based model to detect fraudulent ad comments and metadata with 89% classification accuracy.
• Deployed a Flask-based REST API for real-time fraud probability scoring and label prediction.
• Incorporated explainable AI (SHAP) for transparent model interpretation, enabling regulatory audit compliance.

Ad Fraud Trend Dashboard & Visualization Tool

• Designed and deployed a Streamlit-powered dashboard visualizing fraud trends across geo, device, and publisher segments.
• Leveraged Matplotlib, Plotly, and Pandas for interactive insights from 250k+ ad event records.
• Containerized using Docker and deployed on AWS EC2, ensuring scalable, secure, and low-latency access.