Suggestions
Matt Rubashkin
Director of ML & DE at Cerebral
Matt Rubashkin is a distinguished professional specializing in leading Machine Learning (ML) and Data Engineering teams at Cerebral, a company focused on expanding access to top-notch long-term mental health care.
With an impressive educational background, Matt pursued a Doctor of Philosophy (PhD) across two prestigious institutions, the University of California, Berkeley, and the University of California, San Francisco. His foundational education includes a Bachelor of Science degree from The Johns Hopkins University.
Throughout his career, Matt has held pivotal roles in various organizations, showcasing his expertise and leadership capabilities. He currently serves as the Director of ML and Data Engineering at Cerebral. Prior to this, he contributed significantly as the Director of Engineering and formerly as the Head of Machine Learning at Ready. Matt also led AI initiatives at Insight Data Science and founded LexiconAI. His journey in the industry began as a Data Engineer at Silicon Valley Data Science.
Matt Rubashkin's academic achievements and extensive experience across multiple domains position him as a well-rounded professional with a deep understanding of ML, data engineering, and AI technologies.
Highlights
Releasing our first results @ Future House: WikiCrow, an automated system for synthesizing scientific knowledge. As a demo, we generated an article for every gene in the human genome that previously lacked a complete Wikipedia article: 15,616 in all! https://lnkd.in/gMb7mqJj
WikiCrow 📚 🐦 creates articles in 8 minutes, is more consistent than human editors at citing its sources, although worse than humans at making correct inferences from sources (99.1% human vs 91.0% WikiCrow); a number that we expect to improve as we mature our systems. This was an incredibly difficult challenge, drawing on 14,819,358 pages from 871,000 scientific papers. Sans WikiCrow, we estimate this would have taken a human expert ~60,000 hour total (6.8 working years).
LLM RAG agents are on the rise, with systems like Elicit, Perplexity, and others. However, WikiCrow stands out for its scale and accuracy. PaperQA, the agent underlying WikiCrow, matches humans on a new, very hard RAG benchmark we developed, and is SOTA on a blinded PubMedQA evaluation. Kudos to the team that made this possible Sam Cox, Jakub Lála, Jon Laurent, Michael Hammerling, Andrew White, Samuel G. Rodriques
We hope that WikiCrow + PaperQA will be foundational tools for the automated AI tools we plan to build in the coming years, and will help us to democratize access to scientific research 🎉
