I have developed and deployed 5+ models into production for large-scale Intelligent Document Processing, leveraging Deep Learning Models including BERT, Masked Auto Encoder, ViT, MosaicBERT, Llama3.1 and Mistralv0.3, automating 40% of 100 million previously manually processed documents.I led designing and implementing a proprietary multi-node distributed training platform, enabling efficient orchestration and fault tolerance for LLM fine-tuning and pretraining. This included continued pretraining Llama-3-8b on 1.1B financial tokens and fine-tuning Llama3 models using the Hugging Face PEFT framework, FSDP, and Triton-based custom fused kernels. These efforts significantly reduced training time and resource usage and achieved a 4X increase in training efficiency.One of my key achievements was spearheading the development of 3 prototypes and MVPs for Retrieval Augment Generation using AWS and Azure. These efforts led to a platform that consistently achieves over 95% accuracy in retrieval performance, making it the go-to solution for prototyping cutting-edge demonstrations like Self-Reflective RAG and Multi-Agent Report Generation.In addition, I developed advanced orchestration techniques using Kubernetes and Axolotl to manage multi-node training setups, ensuring seamless communication between nodes and automated recovery from node failures. By implementing checkpointing and automated recovery processes, I minimized downtime and preserved training progress.To support rapid development and deployment, I designed and modernized an MLOps framework that automates the deployment of various Data Science applications, including document classification and extraction. This modernization reduced deployment time by 50% and allowed other Data Scientists to perform multiple experiments simultaneously using the Hydra framework.

Spearheaded the development and implementation of fairness-aware algorithms, decreasing bias in automated decision-making especially using Deep Learning Models by 25%, enhancing ethical standards across 10 institutions, and improving decision accuracy by 20%through comprehensive data analysis and cross-disciplinary collaboration.Published a comprehensive paper in the British Actuarial Journal titled From Bias to Black Boxes : Understanding and Managing the Risks of AI, influencing industry standards and practices in managing AI risks.

Secured a Köln Fortune Research Grant of €120,000 for developing Automated Breast Cancer Screening technology.Supervised four master's theses and collaborated with two doctors on their PhD theses, working on advancements in Statistically Robust Machine Learning.Developed anomaly detection methods in multi-parametric MRIs using Deep Convolutional Neural Networks with FDR control, enhancing detection accuracy by 25%. Built models using PyTorch including custom kernels in C++/Cuda. Conducted Non-linear Independent Component Analysis using Random Fourier Features, improving data separation quality by 15%.

Developed a Discrete Compound Process model for single-cell modeling, incorporating a novel cost function with regularization, improving parameter estimation consistency in under-sampled regimes by 20%.Automated breast cancer screening using multiparametric MRI and Deep Convolutional Neural Networks, enhancing early detection accuracy by 30%.Enhanced interpretability of Deep Bayesian Convolutional Networks, ensuring invariance to rotations in 3D imaging, leading to a 25% increase in diagnostic reliability.Implemented model selection techniques in Deep Neural Networks, controlling false discoveries of features and improving predictive model reliability by 15%.

Designed and analyzed algorithms to control false discoveries, developing machine learning techniques to manage generalization errors. Achieved state-of-the-art results in Genome-Wide Association Studies (GWAS) for breast cancer, reducing false discoveries by 25%.Developed an efficient sampling algorithm to sparsify a kernel matrix with bounded error in O (n log n) time, improving computational efficiency by 50% over the standard O n^2 complexity.Facilitated efficient implementations of Gaussian process regression and kernel-based hypothesis testing algorithms for large datasets, reducing processing time by 40%.

Work in the field of signal processing, wireless communications and machine learning. Applied variational bayes techniques to turbo coding algorithms.

Design and Analysis of Unsupervised Learning Algorithms, prediction of time series. Worked on a project with EDF (Électricité de France) to predict on a daily/weekly basis the consumption patterns of their customers (tens of millions all over Europe). It involved working with very huge (100's of gigabytes) data sets.

Implemented EM algorithm in C to exploit multi-core architectures and provided API for parallel computing.