AI research initiatives, production systems, and startup ventures
Leading BMW's voice-first LLM computer use agent that operates in-car graphical user interfaces like a human, creating a new layer of AI interaction. Through natural conversation, such an agent enables customers to control and use thousands of apps/services worldwide that lack APIs or integration pathways.
Building a consumer AI-native health and fitness platform for personalized conversational coaching based on wearable data with AI assistant integrations, watch integrations, and sport-specific coaches. From running, triathlon, cycling, and gym training to preparation for fitness and endurance events.
Led customer ML use case from concept to production serving entire China market as technical lead for team of 6 engineers on-site in China. Built production ML system using customer vehicle data with drift monitoring and continuous training. Established global production AI platform fostering cross-regional collaboration across BMW.
Developed patent-pending ML classification model for proactive in-car function recommendations and published the research. Built pipelines processing terabytes of customer vehicle fleet data, created dataset of 716K function activations, and designed proxy labeling methodology to create experience labels from driving context to train predictive models.
Selected for BMW China Accelerator out of 152 teams. Led product discovery for personalized LLM-based point of interest (POI) recommendations through customer interviews, surveys, and PoC development, pitched to top management, and secured exit partner.
Research ML platform enabling transfer of AI infrastructure innovations to production ML platforms. Built offboard and onboard (cloud and in-car) features improving how AI at BMW is built, deployed, and operated, enabling delivery of proactive, personalized, and context-aware AI functions to customers.
Built FL infrastructure enabling privacy-preserving machine learning across distributed data sources (vehicle fleet). Critical for training models on sensitive vehicle and customer data while maintaining strict privacy and compliance requirements, as part of BMW's Machine Learning Platform for Research.
Led data science process in an interdisciplinary team of 5 master's students, developing data-driven methodology for Amazon Prime Video to systematically identify high-performing competitor content. Designed content and user acquisition strategy using 4 data sources: user survey and 3 quantitative datasets collected through continuous web scraping (weekly) from multiple platforms to track performance trends.
Developed data pipelines processing telemetry from 25+ autonomous mobile robots at Magazino, a robotics startup with one of Europe's largest advanced robotics teams (acquired by Jungheinrich in 2023). These pipelines drove customer-facing analytics for multiple clients, including 3 multi-billion-euro e-commerce and logistics companies.