Maximilian-David Rumpf

Maximilian-David Rumpf, also known as Max Rumpf, is co-founder and chief executive of SID.ai, an AI research lab training models to find the information other AI systems need. He is a co-creator of SID-1, an agentic retrieval model that learns to search, inspect results, and revise its queries before returning a ranked list of documents.

From accelerator research to learned search

Before founding SID, Rumpf researched AI accelerator design at ETH Zürich. He subsequently co-founded SID with Lotte Seifert; the company was founded in 2023 and joined Y Combinator’s Summer 2023 batch. His work shifted from the hardware that supports AI computation to another constraint on useful AI: finding the context a model needs to answer a question.

SID’s team operates in San Francisco and Zürich. On his personal website, Rumpf calls leading SID the last job he plans to hold. The lab’s research centers on training models to retrieve and reason over information across data sources, supplying larger AI systems with relevant material.

SID introduced SID-1 in December 2025. In May 2026, Rumpf and SID researcher Sam Dauncey published a joint account of its training methods and infrastructure, identifying themselves as the model’s co-creators. Their account connects learned search behavior to a practical engineering problem: thousands of simultaneous training attempts can generate search traffic faster than the backend can serve it, leaving expensive GPUs waiting.

How SID-1 learns to find information

Rumpf and Dauncey argue that fixed retrieval pipelines commit to too many decisions before seeing a question. Query rewriting, search, and reranking can improve results, but a predetermined sequence has limited ability to recover when its first approach fails. SID-1 instead works across multiple turns, choosing search tools, reading returned content and metadata, and deciding whether to search again or report its results.

• Reward retrieval, ranking, and speed. SID trains the model on questions paired with the documents needed to answer them. In the training setup Rumpf and Dauncey describe, each batch contains 256 questions, with 16 attempts per question. A modified version of GRPO, a reinforcement-learning algorithm introduced by DeepSeek, compares the attempts and steers the model toward those that find the right documents, rank them well, and finish quickly. The model learns the search process as well as the final selection.

• Let tool choice emerge through training. SID-1 can choose vector search, BM25 keyword search, and metadata filtering. The team observed it increasingly favoring vector search while continuing to use keyword search for some tasks. It also learned to draft a plausible answer document and use that document’s embedding as a search query—a technique called hypothetical document embeddings, or HyDE. This can place the query closer to relevant answer documents in embedding space.

• Search broadly and narrowly in parallel. The team’s example asks for the age gap between the creator of a television soap and a prime minister who represented the creator’s hometown. Solving it requires connecting several documents rather than matching one phrase. SID-1 may issue a narrow query about the soap’s launch alongside broader queries about British television channels. Rewarding speed encouraged the model to make four to eight searches per turn, gathering more evidence without adding a separate reasoning turn for every lookup.

These behaviors are observations reported by the SID team. They illustrate its research direction: train a model to choose and combine search strategies, then evaluate whether those choices find the documents a question requires.

Making repeated search practical

Learned parallelism also increased the demands on SID’s training infrastructure. With 4,096 attempts running in a synchronous training step and roughly 20 searches per attempt in later stages, a step could involve about 82,000 searches. Initial requests arrived in bursts exceeding 1,000 searches per second, sometimes over corpora containing more than 10 million records. Slow retrieval delayed the entire training step.

The team migrated its search backend to turbopuffer. Its separation of query compute from object storage allowed bursty reads to spread across shared query capacity without first copying a complete index onto each new machine. It also supplied multiple search tools through one backend. For research, corpus branching addressed another problem: the team could update a collection while preserving the earlier version that existing training questions depended on. Rumpf and Dauncey describe these capabilities as ways to spend more of their effort improving the model rather than maintaining the search infrastructure.

SID-1’s role as a retrieval subagent applies a similar division of work at inference time. It explores candidate documents and passes selected results to a larger model, reducing the irrelevant documents and intermediate search reasoning that enter the larger model’s context. Returning a ranked document list also lets it fit into retrieval systems that already select the top results.

SID’s published evaluation reports higher recall than embedding-only retrieval and competitive results against larger models at lower latency and cost. Those comparisons are the lab’s own benchmark results, rather than independent validation. Rumpf’s work with SID and Dauncey brings together the research and engineering required to make retrieval trainable: learning how to find useful context and supporting enough repeated search to improve that behavior.