Abadir Yimam

Contributed to demand forecasting for Home Depot's supplier network to optimize production capacity and reduce inventory imbalances by predicting demand 1–4 weeks ahead. This enabled suppliers to allocate spare capacity elsewhere while maintaining Home Depot's service levels.

My key contribution was implementing a multi-model forecasting approach, testing LightGBM and LSTMs before selecting Facebook Prophet for its superior seasonality handling and ability to incorporate business regressors. I engineered features from previous orders, future commitments, inventory levels, lead times, and days of inventory.

Tools and techniques included Python (Prophet, LightGBM, TensorFlow), feature engineering, cross-validation across multiple horizons, error decomposition, and a robust model evaluation framework.

As a result, suppliers could proactively reallocate surplus capacity while maintaining 99%+ fulfillment rates for Home Depot. The Prophet-based solution provided interpretable, reliable forecasts that stakeholders trusted for capacity planning.

I led the redesign of a logistics accident prediction system used to identify high-risk operational scenarios. I re-engineered features, modeling logic, and evaluation strategy to improve both predictive performance and business interpretability.

My key contribution was to rebuild regression and classification models with improved feature engineering and domain-informed logic, increasing model precision by around 40% and sensitivity to approximately 86%, significantly improving early risk detection. I added explainability layers (feature importance and scenario-level reasoning) to support operational and safety decision-making, validated models against historical incidents, and deployed results for stakeholder consumption.

The tools and techniques I leveraged included Python, Scikit-learn, Pandas, feature engineering, model evaluation, explainable ML, and operational analytics. As a result of my efforts, I enabled safety and operations teams to proactively mitigate risk using interpretable, data-driven insights rather than black-box predictions.

Led a comprehensive overhaul of the company's anomaly detection system, replacing an outdated, library-dependent workflow (ADTK) with a fully transparent, custom-built statistical detection framework. The legacy system relied on an unmaintained Python library; I reverse-engineered its core logic and rebuilt it in-house with enhanced customization, robustness, and explainability.

My key contribution was designing and implementing a modular anomaly detection engine with configurable trend windows, user-selectable aggregation methods (mean/median/IQR), and directional anomaly flags. I unified scoring logic, exposed detailed diagnostic outputs (rolling values, deviations, fallback spreads), and strengthened robustness through multi-tier fallback strategies—enabling stakeholders to understand why anomalies were flagged.

The tools and techniques I leveraged included Python (Pandas, NumPy), statistical process control methods, time-series decomposition, and modular software design principles. I also implemented comprehensive logging, validation checks, and interpretability features to ensure transparency and maintainability.