From Inference to Inquiry

Motivation

As non-native Bostonians, we were shocked by how many pairs of shoes and types of jackets we needed to buy when we arrived. When leaving the house, it's common to pack a backpack with up to three outfits when the weather is especially volatile! We think this is an apt metaphor for the expected direction of human-agent interactions: Today's AI models are not specific or current enough to support the needs of rigorous, constantly evolving research initiatives. We're hoping to change that.

When you step into the known, tundra will always be by your side.

Brief history

In early February, Sidharth was asked by a former Dewpoint colleague to improve Claude's ability to surface therapeutically-relevant insights. He rebuilt his colleague's research pipeline in Claude Code, and noticed significant improvements upon implementing Anthropic's Agent Teams feature. But the excitement didn't last — he experienced the first of many instances of Anthropic unexpectedly decreasing token usage limits, and paid handily as a result. It was obvious from that moment onwards that building sophisticated research pipelines on third-party harnesses like Claude Code and Codex would not be sustainable. Beyond improvements in costs, it was clear that self-hosting the entire architecture would afford him the ability to tailor this system to meet the specific needs of expert scientists he knew. But he couldn't do this alone.

In early March, Sidharth asked his long-time tennis partner and close friend Reed what he thought of this plan. Excited by the opportunity to build a product for a real representative customer, and to implement the SWE and research-related background he had been cultivating at Amazon and in his Masters Program at Columbia, Reed agreed to join. The two of them set out to rebuild a fully self-hosted end-to-end 'drug research pipeline' that Sidharth's colleague had requested help to build. As this phase of the work began, they depended heavily on popular 'AI search' APIs for source finding, and agents SDKs, to architect a bloated, inefficient, albeit 'fully independent of the big labs' research pipeline. They were proud that their work was now technologically decoupled from the big labs, but realized that the performance of their pipeline was far from where it needed to be.

By early May, their pipeline had improved, but it was clear that for the time that it was taking to run the end-to-end pipeline, the quality would have to improve substantially for it to be adopted. With these questions and concerns in mind, Sidharth started to reach out to hedge fund analysts and investment bankers to ask about their willingness to use their pipeline. As conversations progressed, it became obvious that there was a more pressing need for them to solve: Scientific source finding. From Sidharth and Reed's previous experiences (roughly a year before starting tundra intelligence) scraping and organizing literature, they realized that there was an obvious next step for them to take. The pipeline was immediately split into 1) Source finding and 2) Analyses. For their users, this meant that they would offer two products: 1) Domain-specific databases and 2) Domain-specific models.

Looking ahead

We are excited by the early progress achieved in the field of 'small models' research. In short, labs in academia and industry are developing new methods to optimize open-weight models with substantially fewer parameters than Opus 4.8 and GPT 5.5 to perform exceptionally well in clearly specified domains and task types. We believe that our early product offering of domain-specific databases perfectly complements our ongoing research and soon-to-be distributed domain-specific models for researchers with specific interests. We are currently engaging in research collaborations with industry partners, and excited to help academic labs and classrooms take advantage of the technology we are developing. If you're a researcher who finds what we're working on to be interesting, please reach out!