Maximilian-David Rumpf, also known as Max Rumpf, is cofounder and CEO of SID.ai, a research company training models to retrieve information. He and Sam Dauncey identify themselves as co-creators of SID-1, an iterative search model.[2][4]
Start with Training SID-1 to beat GPT-5 at search with 1k+ QPS RL →
Links
SID.ai Founder record
X · maxrumpf Cofounder and CEO · SID.ai
Career
Y Combinator’s founder record describes Rumpf’s earlier research on AI accelerator design at ETH Zürich. SID joined its Summer 2023 batch. The company’s focus is finding context for other AI systems by training models to retrieve and reason over data sources.[2]
Contributions
SID-1 repeatedly queries a search backend, reads documents and metadata, then changes its search strategy before returning ranked documents. This differs from a fixed retrieve-and-rerank pipeline that cannot seek new candidates after inspecting poor results. Rumpf and Dauncey describe reinforcement learning that rewards successful retrieval and efficient behavior. Their training system generates many simultaneous searches; moving the backend to turbopuffer addressed search latency that left training GPUs waiting. The reported quality and cost comparisons are SID’s evaluations under the described setup.[4]
Tools, writing & talks
Writing, research & talks
Start here
Training SID-1 to beat GPT-5 at search with 1k+ QPS RLturbopuffer.comA jointly authored account of iterative retrieval, reinforcement-learning rewards and search infrastructure.
Read the fixed-pipeline comparison, then the model’s search loop. Familiarity with retrieval and reinforcement learning helps; the infrastructure discussion explains why training needs a fast search backend.
Coauthor with Sam Dauncey[4]Maximilian-David Rumpf argues classical multi-stage retrieval can't fix agent search because every decision is frozen at design time, and presents RL-trained search results showing meaningful gains over reranking pipelines on accuracy and cost.
Catalog workOriginal source