Julius Lipp
AI & Search
Brief
My journey as a Founder started when a surfing accident left me in bed for a few months. I had way too much time to think and realized I really want to build something that matters. When I got to Google, I discovered how fun software engineering can be—especially when you're on a team that's genuinely passionate about what they do. Working on LLMs in their early days (that same team is now building Google Gemini) gave me a glimpse of AI's massive potential and shaped everything that followed.
I co-founded Mixedbread to build the next generation of search. Our open-source models reached over 60 million downloads on Hugging Face—about 5.15% of all HF downloads in 2024! I'm incredibly grateful to the brilliant team who made this possible. Their talent and dedication showed that a small, focused group can compete with the world's top AI labs.
I'm fascinated by the gap between what LLMs can already do and what we're actually building with them. While the models keep advancing, most applications barely scratch the surface of their potential. I'm excited to explore creative new ways to apply them and see how software evolves as we push these boundaries.
Timeline
Software Engineer @ crossnative
Built modern web applications
Dishwasher @ various restaurants
Washed dishes and kept the kitchen clean
Open Source & Contributions
ProRank: Prompt Warmup via Reinforcement Learning for Small Language Models Reranking
Jun 4, 2025Two-stage approach combining GRPO with fine-grained logit aggregation, enabling SMLs to outperform much larger models at document reranking.
The Hidden Ceiling: How OCR Quality Limits RAG Performance
May 14, 2025Blog post demonstrating how OCR errors create a performance ceiling for RAG systems, and how multimodal models can improve retrieval by a good margin.
The Hidden Ceiling: How OCR Quality Limits RAG Performance
May 14, 2025Blog post demonstrating how OCR errors create a performance ceiling for RAG systems, and how multimodal models can improve retrieval by a good margin.
ProRank: Prompt Warmup via Reinforcement Learning for Small Language Models Reranking
Jun 4, 2025Two-stage approach combining GRPO with fine-grained logit aggregation, enabling SMLs to outperform much larger models at document reranking.