Jamal Mohammed

Time Series Forecasting Platform

Developed a comprehensive ML platform to quantify and benchmark the predictability limits of univariate time series data. The system implements state-of-the-art deep learning architectures including Transformers, N-BEATS, and N-HiTS, providing a rigorous framework for evaluating forecasting model performance against theoretical bounds.

The platform enables researchers and practitioners to understand what can realistically be predicted from their time series data before investing in complex model development. By establishing information-theoretic upper bounds, organizations can make informed decisions about model selection and resource allocation.

Multi-Horizon Forecasting System

Designed and implemented an end-to-end framework for multi-step time series forecasting that combines deep learning architectures with statistical models. The system addresses the challenging problem of predicting multiple future time steps simultaneously, which is critical for applications like energy demand forecasting, inventory management, and resource planning.

The framework incorporates novel techniques for quantifying prediction uncertainty at different forecast horizons, enabling users to understand confidence levels for short-term versus long-term predictions. This multi-horizon approach provides actionable insights for decision-makers operating under different time constraints.

Efficient Subsequence Matching Tool

Engineered high-performance algorithms for fast subsequence matching in large-scale time series data, enabling real-time pattern detection and anomaly identification in industrial systems. The tool implements optimized data structures and approximation techniques to achieve sub-linear search complexity for pattern matching queries.

This system is particularly valuable for continuous monitoring applications where identifying recurring patterns or detecting anomalies must happen in real-time. The algorithms are designed to handle high-frequency data streams common in IoT sensor networks and industrial control systems.

Predictive Maintenance for Industrial Printers

Collaborated with Durst Phototechnik AG to develop and deploy ML-based predictive maintenance solutions for industrial printing systems. The project involved building end-to-end data pipelines from sensor data collection to model deployment in production environments.

Developed deep learning models (LSTM-based architectures) for anomaly detection in multivariate time series sensor data streams. The system monitors multiple sensor readings simultaneously (temperature, pressure, vibration, ink flow) to detect early warning signs of equipment failure, enabling proactive maintenance scheduling.

Worked closely with domain experts and engineers to translate research prototypes into production-ready solutions, including model optimization for edge deployment and integration with existing monitoring systems.