Senior AI Engineer
shipping production AI for enterprises.

Recommendation system for book chapters and journal articles based on past user interactions. Aggregates interactions into a query vector and performs approximate nearest-neighbour search over a million content embeddings stored in a Milvus vector database in real time. Embeddings generated with Mistral.

• Milvus
• Mistral
• Vector Search
• Media

Agentic chatbot automating conversations between customers and the dealership. Draws on internal documents and Snowflake data to provide fact-checked answers, with guardrails aligned to company guidelines for high-quality, compliant, risk-free responses. Deployed on Azure.

• Azure
• Agents
• Snowflake
• Guardrails

Asynchronous, fully automated pipeline translating documents from 137 countries across 40 languages into German — processing 5,000 documents monthly. Integrates a fine-tuned YOLO model for object detection and Chain-of-Thought reasoning with OpenAI models. Deployed on Azure with Function Apps, Service Bus and Blob Storage.

• Azure
• YOLO
• OpenAI
• CoT
• Function Apps

Extended a data-driven sales support tool with LLM-based text analysis for two oncological indications. Collected data via web scraping, evaluated additional data sources, and delivered an MVP on AWS as a stable, scalable pipeline — successfully handed over to the internal data science team.

• AWS
• LLMs
• Web Scraping
• Pharma

Designed and implemented a complex AI backend for real estate appraisal — processing varied unstructured data into high-quality expert reports. Covered image captioning, speech-to-text, LLM generation, information extraction and OCR. Serverless on AWS with Terraform IaC, a GraphQL API, and full MLOps / CI-CD best practices. Live product: velaro.reguvis.de.

• AWS
• Terraform
• GraphQL
• OCR
• Serverless

Sentiment analysis of Twitch comments for an advertising analytics provider, enabling brand-perception measurement during sponsor campaigns. Implemented on GCP with PyTorch and Terraform.

• GCP
• PyTorch
• Terraform
• Sentiment

GDPR-compliant deployment of an open-source LLM on AWS for a bank. Delivered a Retrieval-Augmented Generation backend and Chat UI, provisioned with Terraform inside a private VPC with a VPN server on AWS.

• AWS
• RAG
• GDPR
• VPC
• Terraform

Research and prototyping of an instruction-based German LLM with RLHF, capable of understanding legal cases, identifying applicable legal contexts, and autonomously asking questions to obtain claim-relevant information.

• RLHF
• German LLM
• LegalTech

Built multiple Information Retrieval pipelines that use a vectorized knowledge database to enrich LLM prompts with contextual information — enabling personalized, fact-grounded chatbot interactions. Deployed on AWS with Terraform IaC (auth, load balancing, and the rest of the production essentials). Public reference: paperqa-web-app.

• AWS
• Terraform
• Vector DB
• RAG

Trained, evaluated, deployed and monitored information-extraction models — NER on client, court and insurance correspondence using BERT and a BiLSTM-CRF ensemble. Also built an OCR model to extract information from customer images with a YOLOv5 + TrOCR ensemble. Experiments on Sagemaker, tracked in MLflow.

• BERT
• BiLSTM-CRF
• YOLOv5
• TrOCR
• Sagemaker
• MLflow

Trained, evaluated and deployed document classification models for a SaaS platform handling 500,000+ annual documents from various sources. Using SWEM and transformer-based approaches, achieved 80% automation at 95% accuracy. Experiments on Sagemaker, tracked in MLflow.

• Transformers
• SWEM
• Sagemaker
• MLflow

• Dynamic Time Warping to reconstruct reference time series from real-life data
• Interpretable ML approaches in production
• Automatic Speech Recognition in a banking context
• Bosch Hand Pose Detection Hackathon
• Internal NLP data-preparation pipeline tool
• ML data mart on Amazon Redshift using DBT