Data Scientist & ML Engineer & AI Architect

• Designed, developed and productized several LLM assistants for different internal stakeholder groups including client advisors, health care advisors and the IT department at Sanitas - one of the largest health insurers in Switzerland
• Took over a PoC of an RAG-based LLM assistant for client advisors as a pilot project from the data science team, designed its production architecture and developed it further up to production
• Communicated with all involved stakeholders such as the data science team, devops team as well as the client and healthcare advisors
• Designed and developed a highly object-oriented base RAG-LLM-architecture which allows for an efficient implementation of specific RAG-based LLM assistants
• Integrated RAG-LLM-architecture in the existing micro-service-based IT architecture partially on MS Azure and GCP
• Integration of various LLM models such as Azure's OpenAI GPT-collection and Google's Gemini collection
• Developed the necessary data fabric architecture including data APIs, data pipelines and vector-search data base with PG Vector in GCP - achieved a considerable speed-up in embeddings retrieval compared to the PoC
• Extended and optimized the underlying FastAPI app, vector-search and overall performance of the LLM assistants
• Improved prompt engineering for LLM models such as the Azure's OpenAI GPT-collection and Google's Gemini-collection
• Design of a common base CI/CD Gitlab pipeline as well as specific Gitlab pipelines for all Gitlab projects related to LLM assistants and vector-search
• Integration of the LLM assistants in Teams, internal SanitasGPT as well as Slack
• Tested the LLM assistant for client advisors with a selected team of internal client advisors and integrated the results in the further development
• Supported the devops team during the roll-out of the LLM assistant for client advisors across different locations in Switzerland
• Developed and deployed a devops support LLM assistant for the IT department at Sanitas in Slack
• Developed a PoC for an LLM assistant to support the Sanitas' health advisor team

Technologies: Python, FastAPI, Pydantic, LangChain, Azure AI Services, Azure OpenAI GPT-collection, Google Gemini-collection, Google Vector Search, PG-vector, Vertex AI, further GCP, DBeaver, SQL, Microsoft Bot Restful API, Kybernetes, Docker

• Led the business case study as project manager for extending the in-house cluster manager ParaStation Software Suite to hybrid multi-cloud datacenters with HPC, AI, QC, IoT and business applications in industry
• Worked directly with the C-level suite and the board of ParTec AG
• Represented ParTec AG at the Supercomputing, ML and QC conference ISC 2023 in Hamburg
• Identified potential customers, their current HPC, AI, QC, IoT and business applications, cloud strategies and the resulting requirements for cluster management with ParaStation Software Suite
• Held meetings with potential customers such as the German Weather Services (DWD) and the German Climate Computing Center (DKRZ) as well as with cloud providers such as AWS
• Performed competitor analysis and derived USPs
• Scoped the technical requirements including changes, new feature development, interfaces, packaging and live demo of ParaStation Software Suite
• Scoped the hardware and software requirements for the unique feature of ParaStation Software Suite for high-speed MPI-based HPC, AI and QC computing across on-premise and cloud data centers
• Selected the most suitable cloud services from MS Azure, Google Cloud and AWS regarding data transfer, HPC, AI and QC services
• Wrote the technical specification for the extension of the existing cluster manager ParaStation Software Suite
• Defined related professional services for the extended ParaStation Software Suite

Technologies: ParaStation Software Suite, various cluster managers including NVIDIA Bright, HPE, Atos, etc., HPC Schedulers such as Slurm, Altair PBSPro, etc., Kubernetes, various services from MS Azure, GCP, AWS concerning data transfer, HPC, Ansible Playbooks, etc., distributed communication libs including MPI, Cuda, Gloo and RPC, distributed versions of Tensorflow and PyTorch, AI-accelerated numerical simulation tools such as NVIDIA SimNet, commercial multi-physics solvers such as Comsol, Ansys, etc., visualization software such as Paraview, Catalyst, VisIt, ADIOS2, various QC libs such as in-house HPC quantum integration software QBridge, Qiskit, etc.

• Improved the performance of an existing scenario planing tool based on Catboost and a selection of regression methods for extrapolation by more than 50% to a MAPE of 5-15%
• Improved the performance of an existing daily demand forecast for over 70,000 products based on AWS DeepAR as well as further forecast and extrapolation methods for its features by more than 40%
• Improved the performance of an existing weekly demand forecast for over 70,000 products based on AWS DeepAR as well as further forecast and extrapolation methods for its features by more than 30%
• Developed a new weekly demand forecast for over 70,000 products based on AWS DeepAR as well as further forecast and extrapolation methods for its features
• Introduced software engineering best practices such as object-oriented programming, logging, exception handling, testing and monitoring
• Implemented best practices together with ML Engineer

Technologies: Python 3, AWS DeepAR, Google Fusion Transformer, statsmodels, Catboost, TensorFlow, TensorFlow Recommenders, MlFlow, PyTest, Fast-API, Github Workflows, Databricks, Delta Lakehouse, Terraform, AWS SageMaker and related AWS services, Apache Airflow, Snowflake, Exasol, Power BI

• Design and evaluation of the business case, communication and workshops with all involved stakeholder groups
• Design individual machine learning use cases
• Choice of suitable machine learning algorithms
• Consulting on potential ethical issues as well as their avoidance or mitigation also from a technical perspective
• Evaluation of suitability of Google Cloud and Google Vertex AI architectures for the Peerdom "AI"-component

Technologies: Python 3, TensorFlow, Scikit Learn, XGBoost, Google Cloud, Google Vertex AI, etc.

• Development of the business case together with the managing directors and Head of IT
• Design and development of a Social Media API client and data collection PoC
• Design and development of a deep/machine-learning based trend analysis and forecasting PoC
• Development and evaluation of different image clustering algorithms and dimensional reduction algorithms on top of the deep learning architecture VGG16 and further hand-crafted feature vectors
• Development and evaluation of different indoor scene segmentation and floor detection approaches based on the latest object detection, semantic segmentation, instance segmentation and panoptic segmentation deep learning algorithms for 2D images
• Sentiment analysis of social media posts and comments in different European languages
• Design of several alternative AWS cloud architectures based on AWS RDS with PostgreSQL, AWS CodePipeline, AWS Glue, AWS SageMaker and AWS Deep Learning AMIs
• Cost estimation and technical design of the overall solution
• Coaching of the Head of IT concerning machine-learning and AWS cloud architectures

Technologies: VGG16, Yolo, Python 3, TensorFlow, PyTorch, Scikit Learn, NumPy, SciPy, Pandas, Pytest, Fast-API, Tableau, Matplotlib, pySpark, OpenCV, Social Media APIs (Instagram, Facebook, Pinterest, Twitter), ElasticSearch, AWS RDS, AWS Pipeline, AWS Glue, AWS SageMaker, Nvidia-Docker, CUDA, Nccl, Linux, Git

• Application of various sophisticated techniques to analyse and decompose retail price multi-variate time series for different markets and demand groups
• Creating various visuals to present the customer the results of the analysis and insights for price optimization
• One-step-ahead and multi-horizon forecasting PoC for multi-variate time series based on XGBoost models and several deep-learning architectures (Stacked LSTMS, Wave-net, Transformers) depending on performance for different markets

Technologies: Python 3, XGBoost, TensorFlow, Scikit-learn, Pandas, NumPy, SciPy, Hyperopt, Statsmodels, Matplotlib, Jupyter, PySpark, PyTest, Linux, Git

• Design and development of a Python-software for exchanging data with Avaya (based on a specific protocol) for data analysis of the obtained communication data from Avaya

Technologies: Python 3 (and various libraries), pySpark, Avaya-Server, Linux, Windows, Git

Traveling Europe and Swiss Mountains

• Pre-evaluated C++- and a C#-energy trading software for a combined OpenMP and MultiGPU parallelization
• Consulted the team on which GPU-hardware to buy for desktop computers and on-premise servers
• Coached the team in optimization and parallelization best practices of linear programming, optimization algorithms, machine learning algorithms and linear algebra routines for energy trading
• Coached the team in designing architectures for highly optimized and parallel energy trading software
• Introduced the design and planning to various stakeholders at Axpo and successfully secured internal funding
• Complete refactoring, improved hyperparameterization and optimization of the single-threaded C++-code and the OpenMP-multithreaded C++-code
• Design and implementation of a CUDA-parallelization of the C++-software for multiple GPUs
• Design and implementation of a combined MultiGPU + OpenMP-multithreaded version of the C++-software
• Design and implementation CUDA-parallelization of an optimization algorithm in the C#-energy trading software based on CUDA managed
• Unit, integration and system tests including memory and speed tests
• Speed evaluation for OpenMP, GPU as well as combined parallelization for various power plants

Technologies: C++, C#, CUDA, CUDA-managed, OpenMP, Posix-Multithreading, Boost, Python, Git, Visual Studio, Windows, Windows-Server, Docker, Jenkins

• Development of a Finite Element Method for Stochastic Navier-Stokes-Equations for meso-scale fluid flow
• Design and implementation of a large-scale massively parallel High-Performance Computing architecture (C++,CUDA,MPI)
• Unit, memory and speed tests
• Setting up a High-Performance Computing (HPC) Cluster (MPI+GPU parallelism) for large-scale simulations in AWS (EC2) and running simulations
• Award (HPC resources worth 20000 EUR) received for the excellent optimization and parallelization of MesoPhy by Arctur HPC in Feb 2018
• Running simulations on the HPC-Clusters of Arctur Validation of the Finite Element method and the software against experiments and for real-life applications

Technologies: C++, MPI, CUDA, Trilinos, Eigen, Boost, CMake, Git, Jenkins, Docker, Linux, AWS (EC2, S3, Starcluster)

• Technical lead and architect for the first AI-based, adaptive context-aware, continuous authentication & risk-detection component "nevisDetect" for the NEVIS security suite. NEVIS including "nevisDetect" is deployed in many financial institutions and government offices in Switzerland and other countries to prevent sophisticated malware attacks in a highly automated way
• Design of the context-aware authentication tool of nevisDetect for the evaluation of client-side and network layer data based on rules
• Design of the continuous authentication tool of nevisDetect for the evaluation of client-side and network layer data based on several advanced machine/deep-learning and especially anomaly detection algorithms
• Communication, collaboration and projects scoping with various stakeholders including the board, C-suite, customers, malware & network experts as well as IT security scientists from ETH Zurich

Technologies: Python, TensorFlow, pySpark, Java, Apache Spark, Apache Kafka, Apache Storm, various other libraries from the Apache Spark Ecosystem, Couchbase, Cassandra Git, REST-APIs, Linux, Docker