The Role of AI in Research: How Reliant Aims to Automate Menial Tasks
AI models have demonstrated their capabilities across various domains, but what specific tasks should we prioritize for automation? Ideally, these should involve repetitive, time-consuming tasks—such as those commonly found in the fields of research and academia. Reliant is positioning itself to specialize in the tedious data extraction processes that currently burden overworked graduate students and interns.
“The ultimate goal of AI is to enhance the human experience by minimizing mundane work, allowing people to focus on what truly matters,” stated CEO Karl Moritz Hermann. Within the research community—where he and co-founders Marc Bellemare and Richard Schlegel have extensive experience—literature reviews are a prime example of such tiresome tasks.
Every research paper references previous studies, yet locating these sources amid the vast sea of scientific literature can be daunting. Systematic reviews, for instance, often cite thousands of sources. Hermann recalled a study where “the authors had to sift through 3,500 scientific publications, many of which turned out to be irrelevant. It translates into considerable time spent extracting minimal useful information—this is precisely the kind of task we believe AI should automate.”
They realized that contemporary language models could assist in this effort. For instance, an experiment using ChatGPT yielded an 11% error rate in data extraction. While this result is commendable, it falls far short of the accuracy researchers truly require.
“That’s simply not acceptable,” Hermann emphasized. “In knowledge-intensive tasks, even minor errors can have significant consequences.”
Reliant’s flagship product, Tabular, integrates a powerful LLM (Llama 3.1) enhanced by proprietary techniques, making it significantly more effective. In the aforementioned study extraction, their model achieved zero errors. Users can input large volumes of documents, specify the data they need, and Reliant adeptly extracts the desired information, regardless of the format's organization. The extracted data, along with any requested analyses, is then presented through an intuitive user interface, facilitating deeper exploration of individual cases.
“Our users need to interact with all data simultaneously. We are developing features that enable them to edit the data or seamlessly transition from data to literature. Our mission is to guide users in effectively directing their attention,” explained Hermann.
This tailored application of AI—while less glamorous than a digital companion—has the potential to significantly advance scientific research in many technical fields. Investors have recognized this promise, contributing to an $11.3 million seed funding round, led by Tola Capital and Inovia Capital, with participation from angel investor Mike Volpi.
Like many AI technologies, Reliant’s solution requires substantial computational power, prompting the company to invest in its own hardware rather than relying on third-party providers. This in-house approach presents both risks and rewards: the equipment is costly to maintain but allows for deeper exploration of technical challenges.
“One key insight we've gained is that providing accurate answers becomes challenging with time constraints,” Hermann elaborated. For example, when a scientist requires novel data extraction from a hundred studies, they can either accomplish it quickly or thoroughly—but not both, unless they anticipate user inquiries and preemptively compile the information.
Many users share similar questions, which allows Reliant to offer early answers as a baseline. “We can condense extensive texts into more digestible formats. While it may not perfectly match what users seek, it streamlines our work,” remarked Marc Bellemare, the startup’s chief science officer.
Consider this analogy: if tasked with extracting themes from a thousand novels, would you wait for someone to ask about specific character names, or would you proactively collect that information—including details about locations and relationships—knowing it would be in demand? The latter approach, if computational resources allow, is far more efficient.
This proactive extraction enables models to address the inherent ambiguities and nuances found across different scientific disciplines. For instance, a single metric may imply various meanings in fields such as pharmaceuticals, pathology, or clinical trials. Additionally, language models may produce diverse outputs based on how questions are framed, challenging Reliant to transform ambiguity into clarity. “This is achievable only if you are committed to investing in a specific science or domain,” Hermann asserted.
Reliant’s immediate focus is to demonstrate that its technology can be self-sustaining before pursuing more ambitious plans. “To foster meaningful progress, it's essential to have a grand vision alongside a concrete starting point,” Hermann stated. “For sustainability as a startup, we aim to collaborate with for-profit entities, as they provide the funding necessary for our GPU infrastructure while ensuring we operate without losses.”
While one might anticipate competitive pressure from companies such as OpenAI and Anthropic, which are heavily investing in structured data tasks like database management, Bellemare maintains a positive outlook: “We are capitalizing on a growing momentum. Any enhancements to our technology stack benefit us, as the LLM is just one of several advanced machine learning models we’ve developed using proprietary data.”
The transition of the biotech and research industries to AI-driven environments is just beginning and may unfold gradually. Nevertheless, Reliant appears to have established a solid foundation for success.
“As for our approach, if you’re content with a 95% accuracy rate and handle occasional errors, that’s fine,” Hermann concluded. “Our focus is on precision and recall, where accuracy truly matters, and we’re committed to maintaining that standard while allowing others to pursue a different path.”
Ranking
