Mercedes-Benz: AI Podcast Generation
Personalized In-Car News with Generative AI
Problem
In-car infotainment systems often bombard users with irrelevant or generic content. With limited attention during driving, there’s a growing need for personalized, voice-first news experiences that align with user preferences and contextual information such as the location and destination.
Approach
Collaborating with Mercedes-Benz R&D, I designed a prototype that uses large language models (LLMs) to summarize, personalize, and voice-deliver news stories based on user interests. The research included a literature review of personalization techniques and recommender systems, a series of expert interviews on driver behavior and media use, and a user study with 13 participants.
Solution
The final product was a high-fidelity prototype of an in-car news system that used real-time RSS feeds and generative AI to generate concise, personalized news podcasts. It integrated content filtering (by category, source, and sentiment) and supported hands-free, voice-based delivery for safety.
The prototype was built in ProtoPie, with a Python backend using LangChain to orchestrate multiple OpenAI API calls and enable retrieval-augmented generation (RAG) from a connected knowledge base.
Impact:
The first usability tests (conducted in Stuttgart, Germany) showed high interest in the prototype. I AB tested the final design against others and concluded with the one that is seen in the video. I presented the prototype to the leadership team of the R&D department, and the concept is currently further being developed by Mercedes-Benz and soon enters the beta in the cars:
The process of building and testing this prototype was documented in the form of a bachelor thesis that can be downloaded here.
Mercedes-Benz: AI Podcast Generation
Personalized In-Car News with Generative AI
Problem
In-car infotainment systems often bombard users with irrelevant or generic content. With limited attention during driving, there’s a growing need for personalized, voice-first news experiences that align with user preferences and contextual information such as the location and destination.
Approach
Collaborating with Mercedes-Benz R&D, I designed a prototype that uses large language models (LLMs) to summarize, personalize, and voice-deliver news stories based on user interests. The research included a literature review of personalization techniques and recommender systems, a series of expert interviews on driver behavior and media use, and a user study with 13 participants.
Solution
The final product was a high-fidelity prototype of an in-car news system that used real-time RSS feeds and generative AI to generate concise, personalized news podcasts. It integrated content filtering (by category, source, and sentiment) and supported hands-free, voice-based delivery for safety.
The prototype was built in ProtoPie, with a Python backend using LangChain to orchestrate multiple OpenAI API calls and enable retrieval-augmented generation (RAG) from a connected knowledge base.
Impact:
The first usability tests (conducted in Stuttgart, Germany) showed high interest in the prototype. I AB tested the final design against others and concluded with the one that is seen in the video. I presented the prototype to the leadership team of the R&D department, and the concept is currently further being developed by Mercedes-Benz and soon enters the beta in the cars:
The process of building and testing this prototype was documented in the form of a bachelor thesis that can be downloaded here.
Mercedes-Benz: AI Podcast Generation
Personalized In-Car News with Generative AI
Problem
In-car infotainment systems often bombard users with irrelevant or generic content. With limited attention during driving, there’s a growing need for personalized, voice-first news experiences that align with user preferences and contextual information such as the location and destination.
Approach
Collaborating with Mercedes-Benz R&D, I designed a prototype that uses large language models (LLMs) to summarize, personalize, and voice-deliver news stories based on user interests. The research included a literature review of personalization techniques and recommender systems, a series of expert interviews on driver behavior and media use, and a user study with 13 participants.
Solution
The final product was a high-fidelity prototype of an in-car news system that used real-time RSS feeds and generative AI to generate concise, personalized news podcasts. It integrated content filtering (by category, source, and sentiment) and supported hands-free, voice-based delivery for safety.
The prototype was built in ProtoPie, with a Python backend using LangChain to orchestrate multiple OpenAI API calls and enable retrieval-augmented generation (RAG) from a connected knowledge base.
Impact:
The first usability tests (conducted in Stuttgart, Germany) showed high interest in the prototype. I AB tested the final design against others and concluded with the one that is seen in the video. I presented the prototype to the leadership team of the R&D department, and the concept is currently further being developed by Mercedes-Benz and soon enters the beta in the cars:
The process of building and testing this prototype was documented in the form of a bachelor thesis that can be downloaded here.