Machine Learning Intern - Pixalione

🧠 Problem Statement & Motivation

Each client manages one or more Google Ads campaigns with a fixed daily budget. Spending occurs per keyword match and is driven by live auction dynamics, click rates, and impressions. However, most clients lack tools to anticipate daily trends, often resulting in inefficient budget use.

Our goal was to create a predictive system that helps clients understand and control their spending behavior, improving their ROI by proactively adjusting campaign strategies.

🔧 My Contributions

• Built a robust machine learning pipeline to predict daily ad spend using past trends in click-through rates, impressions, and user profiles.
• Enabled per-client budget optimization, integrating live forecasts into business logic.
• Deployed a Flask-based web API to serve predictions and trigger real-time notifications.
• Daily Forecast Email: Clients receive a daily email with the forecast of today’s spending, helping them stay informed and plan accordingly.

🚀 Technology Stack

🧠 ML/DS Tools

• scikit-learn, XGBoost, Facebook Prophet, LightGBM,
• Pandas, NumPy for feature engineering and preprocessing
• Time-series modeling: seasonal decomposition, exponential smoothing (ETS), ARIMA, Prophet
• Deep learning: RNN, LSTM for sequential trend prediction
• Tree-based models: Decision Trees, Random Forest, Gradient Boosted Trees
• Model evaluation: cross-validation

🖥️ DevOps & Backend

• GitLab CI/CD for automation and testing
• Docker for containerized deployment
• Azure Cloud for hosting and data storage
• Flask + REST API for real-time interaction

Outcomes

• Established Forecast Accuracy Baseline: Created the initial baseline for ad spend forecast accuracy, enabling future measurement and improvement efforts.

• Enabled hands-free campaign monitoring for clients with low technical overhead.