Stephan B.-Freelance Data Scientist

Setting up a hybrid ML architecture on AWS and on-premises

Developing custom machine learning models for computer vision and information extraction

Evaluating and prototyping with different agentic AI tools, LLMs, and MCP

Collaborating with product owners to define functional requirements and security aspects

Leading a team of data scientists and data engineers

OCR pipeline: Training an EasyOCR model on domain-specific and synthetic datasets (ID cards, passports, driver’s licenses)

Segmentation model: Implemented a segmentation model in PyTorch for precise cropping of documents and perspective correction

ID classifier: Developed a CNN-based model in PyTorch to classify document types

Hologram detection: Developed a specialized classification model in PyTorch to verify holograms

Converted models to ONNX and TensorFlow Lite, including quantization and pruning of model weights to meet real-time performance requirements

Achieved average inference times of under 200 ms on mobile devices

Over one million successful real-time idents

Reduced manual verification effort by more than 90 %

Tech stack: AWS SageMaker, Bedrock, Rekognition; multimodal LLMs, Pydantic, FastMCP, prompt engineering; PyTorch, PyTorch Lightning, TensorFlow Lite, ONNX; TorchVision, OpenCV, EasyOCR

Streaming ML for Customer Message Processing (Oct 2019 – May 2021)**

Built a real-time ML pipeline for classifying customer communications

Created live dashboards for message flow, predictions, and system monitoring

Designed a VAIT-compliant ML lifecycle with audit trails, Jenkins CI, and deployment on Kubernetes

Tech stack: Kafka, OpenShift, Jenkins, MLflow, Python, scikit-learn, XGBoost, Universal Sentence Encoder

Enabled real-time processing of over 20,000 messages per day

Delivered fully traceable and regulatory-compliant ML workflows

NLP Pipeline for Semantic Search and Entity Linking (Aug 2018 – Sept 2019)**

Developed and implemented an NLP pipeline using BERT-based models for NER and entity disambiguation

Built a semantic search engine with Elasticsearch and Kibana dashboards for query analysis

Linked extracted entities and relationships in a Neo4j knowledge graph, enabling graph-based search and interactive exploration via Neo4j Bloom

Tech stack: PyTorch, BERT, Elasticsearch, Kibana, Neo4j, Bloom, Azure

Provided more timely risk assessment for credit insurance by incorporating current news events

Increased efficiency for claims adjusters by enabling faster identification of relevant historical claims

Sales Funnel Optimization for Loan Products (May 2018 – July 2018)**

Applied clustering and classification models to segment user behavior and identify drop-off patterns in the sales funnel

Performed data analysis, feature engineering, and model development

Developed an interactive web application to visualize conversion paths, customer segments, and model results for business stakeholders

Tech stack: Python, PySpark, Pandas, NumPy, scikit-learn, SQL, Plotly, Dash