loading
loading
Off the Wire Press Releases New multimodal decision model combines domain specialization, calibrated decisions and adaptive verification to bring greater precision to enterprise AI agents SANTA CLARA, Calif., Oct. 28, 2026 — Articul8, a provider of domain-specific generative AI solutions for enterprises, today introduced Bivia, a multi-modal decision model that decides, locates, and proves what it sees and can be tailored to an organization’s industry knowledge, policies, and operating procedures. Built on billions of decisions made within Articul8’s enterprise AI platform, Bivia is now being offered through an API, allowing developers to build AI agents that make decisions across text, images, videos, tables and charts without having to call a generative model for every step. As enterprises put AI agents to work, they face a growing challenge: the cost of using powerful generative models to make thousands of relatively simple decisions. According to McKinsey’s State of AI in 2026 report, 40% of respondents from large enterprises said their organizations were scaling AI agents, up from 27% the previous year. The same research found that approximately one in five respondents reported that AI-related operating costs, including token costs, were constraining their organizations use of AI. Unlike generic decision models, Bivia can be tuned to an organization’s industry, policies, procedures, and historical decisions, allowing it to make judgments that reflect how the business operates. It can also evaluate text alongside engineering drawings, equipment images and charts. “We’ve been using decision models within Articul8 for more than three years because we saw how much unnecessary work enterprises were asking generative models to do,” said Dr. Arun Subramaniyan, Founder and CEO of Articul8. “When an AI agent is working through a complex task, it doesn’t need a powerful generative model to answer every question along the way. And businesses need those decisions to reflect their own knowledge and operating requirements. That’s why we’re making this technology available to everyone.” From Hours of Content to Precise, Evidence-Backed Decisions Point Bivia at an image, document, audio clip or feature-length video and ask a question in any language. It responds immediately, grounds its answer in the exact pixels, words or timestamps supporting it, and provides a confidence score. When the evidence is insufficient, it can abstain rather than produce an unsupported answer. In a live demonstration, Bivia processed every frame of a 52-minute, 4K documentary filmed from the International Space Station (ISS) in a single request. It transformed the footage into a searchable index, providing an at-a-glance map of what appeared on screen throughout the film and allowing users to ask questions and jump directly to the relevant moment. Among the footage were fleeting events lasting just seconds, including a pass over city lights on Earth’s night side and a glimpse of the aurora. Bivia identified these moments because it analyzed every frame rather than sampling selected images, which could have missed them. “Everyone’s models can caption a thumbnail. Bivia watches the entire thing and tells you the three seconds that matter, with the receipts,” said Dr. Subramaniyan. “It’s the difference between a model that talks about your data and one that actually reads all of it.” Bivia’s capabilities include: - One model across multiple modalities. Text, documents, images, video, audio and raw sensor signals are processed through a shared model architecture rather than a pipeline of separate models. - Structured decisions instead of generated text. Eight native decision primitives, decide, yes/no, choose, score, select all, quote the source, find the region and extract the field, return structured, machine-readable answers in a single pass. - Evidence-backed answers. Responses can identify supporting regions within images, exact text passages and character offsets in documents, and timestamps in video, allowing users to trace decisions to their sources. - Calibrated confidence. Each answer includes a confidence score, and the model can abstain when the available evidence is insufficient. - Comprehensive content analysis. Bivia can process entire videos, documents and audio recordings in parallel, turning lengthy content into searchable indexes rather than relying on selected samples. In industrial environments, Bivia can assess technical requirements against supporting documents and images, including identifying discrepancies between maintenance records and equipment photographs. In Articul8’s internal testing, Bivia achieved 100% accuracy on selected energy and hardware design (Verilog) tasks after being trained on domain-specific data. These in-distribution results demonstrate its ability to be specialized for specific industrial tasks, rather than its performance on previously unseen data. Designed for Decisions at Scale Bivia supports many types of decisions including Choice, Boolean, Score, Set and Span. Developers can use these to build applications for evidence verification, tool routing, document processing and task-completion checks. Three operating modes allow applications to make immediate decisions or apply additional computation and verification when required. The model can also abstain when evidence is insufficient, allowing the application to request more information or escalate the decision. In Articul8’s internal workflow completion tests involving tool routing, Bivia scored 0.91 against Jev’s 0.84. Bivia is designed to deliver consistent decisions regardless of how answer options are presented, making it extremely robust to changes and inputs. While Jev and other leading decision models changed their answers in approximately one in eight cases, Bivia returned the same answer every time. Because Bivia makes decisions without generating output tokens, developers can avoid using more expensive generative models for routine tasks, helping to reduce the cost of running AI agents at scale. The model is available through an API, either as a standalone offering or as part of Articul8’s broader enterprise AI platform. Bivia is already producing substantial savings to Articul8’s customers such as Hitachi Energy, CTC, AZing, Megazone and others. Built on Deep Research and Industrial AI Expertise Founded in January 2024 as an Intel spinout, Articul8 is led by Dr. Arun Subramaniyan, an aerospace engineer with a PhD from Purdue University and previous leadership experience at Intel, AWS and GE. The company brings together a global team of AI specialists from industry and academia and reports having filed 68 patent applications. Articul8 completed its Series B financing in June 2026 at a $500 million pre-money valuation, alongside a strategic partnership with a global industrial software company. At the time, Articul8 reported more than $100 million in cumulative contract value, with a platform comprising more than two dozen domain-specific AI models and connections to over 25 enterprise and operational systems. About Articul8 Articul8 AI provides domain-specific generative AI for mission-critical and regulated industries, including energy, manufacturing, aerospace and defense, financial services, and semiconductors. Founded in January 2024, the company helps enterprises turn complex operational and engineering data into reliable, actionable intelligence. Its platform combines multimodal data understanding, domain-specific AI models, and intelligent agent orchestration to deliver accurate, explainable AI applications that work with existing enterprise systems. Articul8 supports deployments across public, private, and sovereign clouds, as well as air-gapped environments. The company is SOC 2 Type II certified, with solutions available through the AWS, Microsoft Azure, Google Cloud, and Databricks marketplaces. For more information, visit articul8.ai. Source: Articul8
document fetched 2026-10-29 03:01 · canonical https://hpcwire.com/off-the-wire/articul8-launches-bivia-to-bring-domain-specific-decision-intelligence-to-enterprise-ai-agents