Rozum Launches AI Reasoning Engine Built For High Stakes Decisions

New reasoning engine born out of a water startup scores 65.7% on Humanity’s Last Exam

SAN MATEO, Calif., March 13, 2026 (GLOBE NEWSWIRE) -- Rozum Corporation today announced the launch of Rozum, an advanced reasoning engine purpose-built for technical leaders making high stakes decisions. Rozum is currently being deployed across bioengineering, finance, health care, and other high stakes industries.

How a Water Desalination Startup Built a Better AI

Rozum’s origins trace back to an unexpected place. The system was first developed inside Waterline Development, a startup working on next-generation desalination technology to address global water scarcity. As the founding team worked through complex engineering tradeoffs for desalination cell design, they relied on large language models for research and technical analysis - and quickly ran into a consistent problem.

Existing AI systems were confidently wrong. They fabricated citations, dropped steps in multi-stage reasoning, and failed silently on the kinds of problems where getting it wrong has real consequences. So the team built a fix: an internal tool that cross-examined multiple frontier AI models simultaneously, verified their outputs, and flagged gaps before surfacing a final answer.

That internal tool became Rozum. In late January 2026, the founders transitioned to building Rozum full time after recognizing its potential far beyond their original use case.

“A lot of AI today is built either to generate quick answers or automate jobs,” said Derek Bednarski, CEO and Founder of Rozum. “But the world’s hardest problems still require careful reasoning. We built Rozum to work alongside experts when the stakes are high.”

Benchmark Performance

Humanity's Last Exam (HLE) is one of the most demanding AI benchmarks available, testing PhD-level reasoning across mathematics, science, law, medicine, and engineering. In internal testing on a stratified random sample of 484 HLE questions, Rozum achieved:

• 65.7% ± 4.8 pp overall accuracy
• +14.6 pp improvement over best performing model within Rozum
• +7.0 pp improvement over highest publicly reported HLE score we are aware of
• Performance gains across all tested scientific domains

Rozum's verification layer, which cross-checks claims across models before generating a final response, flagged unsupported claims in 76.2% of individual frontier model responses and caught source errors in 21.3% of responses. Full benchmark methodology and results are available in the company’s technical white paper.

A New Approach to AI Reasoning

Most AI systems route a question to a single model and return whatever answer it generates. Rozum's patent-pending approach works differently: every query runs across multiple frontier models in parallel, with outputs evaluated, cross-checked, and verified before synthesis into a final response.

The system architecture includes:

• Parallel execution across multiple frontier AI models
• Domain-specific scientific tools, including symbolic mathematics and cheminformatics
• A verification layer that removes unsupported claims, calculation errors, and fabricated citations. Each claim produced by the system is evaluated and classified based on evidence strength before synthesis into a final response.
  

Early Applications

Early customers and pilot partners are already using Rozum to support complex decisions across science, finance, and engineering:

• Investment and infrastructure analysis: Evaluating large capital projects, including energy and real estate investments, and assessing long-term external factors such as power market dynamics
• Bioengineering and R&D prioritization: Helping research teams allocate technical resources, pressure-test experimental paths, and accelerate scientific decision-making
• Scientific and medical research: Supporting researchers working through complex questions in areas such as oncology and interdisciplinary literature synthesis
• Expert analysis and technical writing: Assisting senior leaders with high-stakes writing, structured reasoning, and difficult cross-domain questions

In one pilot, work that had previously required months of literature synthesis and research support was delivered as structured analysis in hours. In a second pilot, a technical team estimated the system saved two months of time during a critical research prioritization process.

Leadership Background

Rozum Corporation is led by CEO Derek Bednarski, who previously spent eight years at Tesla developing machine learning and automation systems for vehicle service operations. Bednarski is also a co-inventor on a patented automation system that improved operational efficiency and customer experience.

Co-founder Eddie Brucculeri brings deep executive finance experience. As Head of Finance at P97 Networks, he built the finance function from the ground up, raised more than $80 million in capital, and helped steer the company through a successful acquisition.

Request Access to Rozum

Rozum is currently available through a limited early access program for organizations working on complex problems in science, engineering, and finance.

Apply for access at getrozum.com

About Rozum Corporation

Rozum Corporation builds reasoning infrastructure for technical professionals in bioengineering, finance, science, real estate, and other technical fields. Founded in 2024 and headquartered in San Mateo, California, the company develops multi-model systems that prioritize factual accuracy and evidence-based verification for high-consequence decisions. Learn more at http://www.getrozum.com/

Contact:
info@getrozum.com

A photo accompanying this announcement is available at https://www.globenewswire.com/NewsRoom/AttachmentNg/a25ede20-9a9d-471a-9986-ff0434bbb4c6

Derek Bednarski, CEO & Founder of Rozum Corporation

Derek Bednarski, CEO & Founder of Rozum Corporation, at the company's headquarters in San Mateo, CA

Legal Disclaimer:

EIN Presswire provides this news content "as is" without warranty of any kind. We do not accept any responsibility or liability for the accuracy, content, images, videos, licenses, completeness, legality, or reliability of the information contained in this article. If you have any complaints or copyright issues related to this article, kindly contact the author above.