Today at our Research@ event, we shared new breakthroughs and tools that will empower global scientists, doctors, developers, and public health officials to advance science and technology. Here are a few highlights: -We’re beginning to test new ways for partners to access Google Earth AI’s agentic capabilities, including integrating our Geospatial Reasoning agent skills directly into custom workflows via an experimental agent-to-agent API. -We created Deep Review, an agentic system that provides rigorous verification across complex knowledge documents. It was originally designed for computer science academic paper analysis, and today, we’re making Deep Review available via API so anyone can use it to improve a variety of technical documents from legal agreements to financial reports. -TimesFM and TabFM are our foundational models engineered specifically to reason over raw numerical distributions. Both models deliver accurate predictions out of the box with zero training, and rank at the top of leading public benchmarks in their fields. They’re now available to all Cloud BigQuery users, and we’ll soon bring these models into Gemini for Business. -MedGemma, our collection of open-weight AI models built on the Gemma models and optimized to understand medical text and images, is pushing the frontier of open health AI with new multimodal capabilities, clinical reasoning, and real-world virtual care insights. Developers are adapting MedGemma in projects as varied as enabling faster triage in overcrowded clinics to screening for eye disease and cervical cancer. Learn more about all of our updates showcased during Research@ → goo.gle/47jJh2Q
Deep Review is the piece I'd want to understand better — an agentic system for rigorous verification across complex knowledge sounds adjacent to LLM-as-judge evals, but the "agentic" framing suggests something more like multi-step verification with tool use. How does its verification loop differ from a standard judge-model setup? The distinction matters for anyone building eval harnesses.
Democratizing frontier models for academic research is the exact pressure-test our ecosystem needs to move past closed-garden iteration cycles. Tying $3M in compute directly to AlphaFold and weather systems creates the baseline leverage European labs require to compete at scale.
TimesFM and TabFM in BigQuery with zero training is the part I'd watch. Most small companies already have years of sales history but never had a forecast without a data team. Helping BigQuery customers try that is work we enjoy at D3V.
The interesting one here is Deep Review. It is noteworthy to ship verification as its own API rather than as a feature inside something else because it is usually the part teams build last and least well. However, the jump from academic papers to financial reports and legal agreements is a big one though, since those documents fail in quite different ways.
What stands out is the shift from AI generating insights to AI supporting real decisions. Making these capabilities accessible across industries could change how complex challenges are approached.
The differentiator is verification at scale. Deep Review becomes operationally valuable when teams can trace each conclusion to evidence and a named reviewer.
Really exciting to see research moving in this direction! 🚀 Having worked on AI/ML research, healthcare data, IoT monitoring, and agentic systems, I can relate to how challenging it is to turn complex data into something actually useful. I'm particularly interested in MedGemma and Google Earth AI. Imagine combining geospatial intelligence, real-time data, and AI agents to solve problems before they become serious! That's the kind of technology I love building. Of course, making AI reliable in the real world is where the actual engineering fun begins! 😄👏
Deep Review stands out to me. Rigorous verification of legal agreements and financial reports through an API could save teams a lot of time. Exciting to see research turn into practical tools.
Interesting to see AI moving into more specialised, real-world applications. Thats where its practical value becomes clearer.
Exciting progress in making advanced AI tools more accessible to researchers and developers. Bringing agentic workflows and specialized models into real-world applications could accelerate scientific discovery.