Projects & Talks
A selection of my projects and talks, with hands-on demos and in-depth resources.
No tag suggestions yet.
Free-text search still filters titles, descriptions, locations, and link labels below.
Talks
Conference talks and community sessions covering quantization, multilingual speech, sustainability, and other hands-on lessons from the AI space.
A joint presentation at AI.SUMMIT Hamburg, this talk offered a case study on real-world GenAI applications in the banking sector. It detailed the development and successful adoption of an in-house, LLM-powered chatbot that connects key corporate knowledge resources and achieved an impressive adoption rate.
In a breakout session at Green AI Day, I discussed with Steffen Röcker from Red Hat how quantizing models can save resources without sacrificing quality, applicable not just to LLMs but also TTS models.
Presented at CNCF GenAI SummerJam, this talk explored how Large Language Models can be engineered to speak with emotion, even in German.
Deep dive on quantization strategies for sustainable inference, presented at CNCF Sustainability Week Stuttgart.
Projects
Models, fine-tunes, and creative tools spanning speech, image, and language work that I actively maintain or recently shipped.
A German finetuned TTS model based on Resemble AI's Chatterbox. It features emotion exaggeration control, voice cloning, and support for various vocal expressions like laughter and sighs. To generate speech with specific emotions, use tags like <hahahaha> or <ughh> in your input text.
A German TTS model family based on Orpheus-3B, available in two versions: 'Natural' for realistic voices and 'Synthetic' for expressive, emotional speech with outburst support. Supports different speakers by using natural names like Jakob or Sophie.
A lightweight German TTS model based on SmolLM2-135M. It includes a GRPO variant to reduce speech hangups, trained on 7,000 hours of audio and optimized for performance.
A collaboration to finetune the LLaSA TTS framework for Italian (Llasagna) and German (Kartoffel), enabling natural-sounding speech generation in new languages.
A German text-to-speech model based on Llasa-1B, trained on 7,000 hours of high-quality audio. It supports both random voice generation and voice cloning from a reference audio sample.
A LoRA model for the FLUX text-to-image model, trained to generate images in a watercolor style.
A LoRA for the FLUX model that creates images in a retro line-drawing style, based on illustrations from the 'Dictionnaire universel d'histoire naturelle'. Use 'TOK' to trigger the style.
A LoRA for the FLUX model designed to generate cute 3D characters and scenes. The style is activated with the trigger word 'NEOCUTE3D'.
A German instruction-tuned LLM based on BLOOM. It was fine-tuned on a translated and filtered WizardLM dataset to enhance its capabilities for various tasks in German, including code generation.
An instruction-tuned German LLM built on MosaicML's MPT-30B. This model was QLoRa-finetuned using a translated WizardLM dataset to improve German-language interactions and code generation.