Dr. Christos Tzagkarakis

Senior Signal Processing & AI Engineer, BreezeLabs (2024–Present): audio-based AI for athletes’ breathing-pattern analysis using deep learning.

AI Quant, Thira LLC (2024–Present): statistical and AI models for financial time-series and real-time trading decisions.

Post-doctoral Researcher, TU Delft / Google collaboration (2015–2017): acoustic source localization research that led to a Google patent.

Full experience details → Full education details →

My research centers on machine learning, signal processing, and data-driven analytics, with a particular emphasis on sparse representation theory, compressed sensing, and dictionary learning.

Foundations: Audio Modeling & Compressed Sensing

My earliest research explored statistical approaches to musical genre classification, using generalized Gaussian and alpha-stable models to capture the non-Gaussian characteristics of audio features. This initial exposure to statistical signal processing shaped much of what followed. From there, I moved into multichannel audio coding for immersive environments, developing sinusoidal models for efficient encoding and transmission of spatial audio signals. This led to contributions in compressed sensing applied to audio, demonstrating that sparsity-aware methods could significantly reduce bitrate requirements while preserving perceptual quality.

Robust Speaker Recognition

I then focused on robust speaker recognition, where the core challenge was achieving reliable identification with limited training data in noisy environments. I introduced sparsity-based classification methods to the speaker identification problem, demonstrating significant improvements over standard statistical baselines when both training and test sessions were short. I then developed a discriminative dictionary learning framework that jointly optimized the learned representation, its discriminative structure, and the classification model, yielding consistent gains across diverse real-world noise conditions. In parallel, I developed a low-rank matrix recovery approach to reconstruct missing or corrupted spectral features, leveraging the inherent redundancy of speech signals to achieve substantially higher identification accuracy than existing imputation methods, particularly at very low signal-to-noise ratios.

Acoustic Source Localization

A significant thread of my research has been acoustic source localization, where I designed sparsity-based algorithms for jointly localizing and recovering wideband acoustic sources using microphone arrays. A key challenge in this setting is that real source positions rarely align with a predefined spatial grid, causing energy leakage and spurious location estimates. To address this, I developed a grid-shift approach that evaluates the acoustic scene across multiple shifted grids, enabling accurate localization and signal acquisition even under noisy and reverberant conditions. This work resulted in a US patent and contributed to the broader research effort on sound source localization in wireless acoustic sensor networks.

Applied AI Across Domains

In more recent years, my academic research broadened into applied AI across several domains. In IoT security, I developed lightweight sparse representation methods for botnet attack detection at the network edge. In social media analysis, I built explainable pipelines for automated bot detection and trend monitoring. In cybersecurity, I designed AI models for threat and vulnerability assessment targeting healthcare IT infrastructures, translating risk indicators into actionable mitigation strategies. In healthcare, I developed predictive analytics pipelines for mortality prediction in critically-ill patients, combining classical ML with NLP-driven analysis of clinical data. I also explored sound event classification using discriminative dictionary learning, short-term time series forecasting under resource-constrained IoT settings, and probabilistic forecasting for electricity price prediction, applying statistical and machine learning techniques to capture the uncertainty and volatility inherent in energy markets.

Applied & Translational Work

Beyond academic research, a part of my work has been driven by real-world deployment in industrial and startup settings. In social media intelligence, I developed deep learning models for user identification through network traffic analysis and mobile device classification based on behavioral communication patterns. I also developed deep learning models for breathing pattern analysis in athletes, transforming earbud microphone data into performance insights using audio embeddings and transformer-based architectures. More recently, I have been building AI-driven quantitative models for financial time series analysis and real-time decision-making.

• “IoT challenges: Sparsity at the rescue.” Invited Talk, IBM Research, Zurich, February 2021. Slides
• Invited Talk, Chrome/YouTube Research Conference. Google, Mountain View, CA, August 2016. Slides
• Poster, Spring School on Sparse Representations and Compressed Sensing. Ilmenau, Germany, April 2016.
• Poster, Google European Doctoral Workshop on Speech Technology. London, UK, April 2014. Poster