The team includes Will (Huichen) Wang, a Ph.D. student at the University of Washington and Tableau intern, Tableau Research Director Vidya Setlur, and Northwestern University Professor Larry Birnbaum.<\/p>\n
Let\u2019s explore the motivation and process behind this research, and how Jupybara helps analysts turn data into clear, persuasive, and actionable stories within the Jupyter Notebook environment.<\/p>\n
From Data Exploration to Result Communication<\/strong><\/p>\n As an AI agent\u2013powered smart assistant, Jupybara supports the full data analysis workflow\u2014from initial exploratory data analysis (EDA) to insight-driven storytelling. It\u2019s implemented as a Jupyter Notebook extension, integrating large language models (LLMs) directly into the analyst\u2019s working environment to enable human-AI collaboration.<\/p>\n (A) On the left, when exploring complex data, Jupybara first generates and displays an analysis plan before outputting code.<\/p>\n (B) On the right, in the generated data story, Jupybara articulates the analysis results with precise language, including an introduction, transitions, and strategy explanations. It helps users gain deeper insights and uses domain knowledge to translate data findings into action plans.<\/p>\n Additionally, based on a three-dimensional design framework, Jupybara integrates data visualization theory, narrative logic, and communication principles. These dimensions guide how Jupybara generates, optimizes, and refines data stories\u2014ultimately converting insights into real-world impact.<\/p>\n 01 Semantic Precision<\/strong><\/p>\n Ensures that data insights are communicated with both accuracy and factual grounding.<\/p>\n For example, when describing a steep decline in average hourly wages, Jupybara might suggest terms like \u201cdecline,\u201d \u201cdecrease,\u201d or \u201cplummet,\u201d each with a distinct semantic tone\u2014showing how language can stay tightly aligned with statistical trends.<\/p>\n 02 Rhetorical Framing<\/strong><\/p>\n Enables persuasive, audience-targeted communication so insights resonate with intended recipients.<\/p>\n For instance, when analyzing gender pay gaps, Jupybara\u2019s story assistant might begin with a headline-style statement: \u201cWomen in full-time roles earn 17% less than men,\u201d followed by explanatory detail: \u201cEven after adjusting for industry and education, the gap persists\u2014indicating systemic bias rather than structural factors.\u201d<\/p>\n 03 Real-World Relevance<\/strong><\/p>\n Connects insights directly to real decision-making contexts, combining industry background with practical action.<\/p>\n For example, upon detecting a significant gender pay gap in full-time roles, Jupybara might recommend: \u201cOrganizations should establish transparent pay systems and conduct biannual gender-based promotion audits.\u201d<\/p>\n Multi-Agent Collaboration for Human-AI Co-analysis<\/strong><\/p>\n Jupybara introduces a multi-agent architecture in which different LLM-based agents specialize in distinct tasks throughout the analysis and storytelling process: some interpret statistical results and visualizations, others check for clarity and persuasiveness, while others ensure alignment with specific industry contexts.<\/p>\n This division of responsibilities is inspired by how human analysts collaborate as a team, enabling iterative refinement of both analysis and storytelling.<\/p>\n Of course, for simpler or time-sensitive tasks, Jupybara also supports a single-agent mode. However, the multi-agent approach opens new possibilities for future research\u2014especially in complex scenarios that require layered reasoning, peer review, or flexible narrative adjustments.<\/p>\n Insights from Cross-Industry Practice<\/strong><\/p>\n To evaluate how analysts interact with AI-assisted tools in exploratory data analysis (EDA) and storytelling, the research team invited nine experienced analysts from finance, healthcare, retail, and other industries to test Jupybara.<\/p>\n Participants used both the single-agent and multi-agent versions to complete an EDA task and generate a data story using a dataset relevant to their domain expertise.<\/p>\n Analysts generally found Jupybara\u2019s \u201cinsight tracker\u201d helpful in reducing the mental load of remembering and organizing the analysis process, keeping thought processes and history more transparent. The \u201cclarification tab\u201d enabled targeted dialogue with the AI to better understand the analysis results it provided.<\/p>\n Additionally, Jupybara proved valuable for refining how analytical conclusions are expressed. Semantic and rhetorical suggestions helped analysts choose better wording and build more persuasive narrative structures.<\/p>\n Jupybara\u2019s suggestions for analytical strategies were also highly practical\u2014such as explaining the rationale behind handling missing values\u2014supporting reflection and explanation of their methodological choices.<\/p>\n Overall, participants consistently agreed that outputs from the multi-agent architecture outperformed single-agent results in terms of semantic precision, rhetorical effectiveness, and real-world relevance.<\/p>\n More Than an Analyst Tool \u2014 A Research Platform<\/strong><\/p>\n Jupybara is not only a powerful tool for improving analyst productivity, but also a research platform for exploring human-AI collaboration. Tableau Research\u2019s evaluations show that Jupybara excels in usability, controllability, explainability, and revisability\u2014key factors in building effective human-AI collaborative systems.<\/p>\n As multi-agent AI tools become more prevalent, Jupybara is just the beginning. In the future, the Tableau Research team may explore agents specialized by domain (e.g., healthcare or finance) or narrative style\u2014or systems that can autonomously adjust analytical strategies based on evolving data.<\/p>\n","protected":false},"excerpt":{"rendered":" In today\u2019s data-driven world, uncovering insights is no longer enough\u2014what truly matters is transforming those insights into actionable outcomes. Whether<\/p>\n","protected":false},"author":92,"featured_media":10319,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"_acf_changed":false,"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":"","_wp_rev_ctl_limit":""},"categories":[1],"tags":[203],"class_list":["post-10318","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tableau","tag-tableau-research"],"acf":[],"aioseo_notices":[],"rttpg_featured_image_url":{"full":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false],"landscape":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false],"portraits":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false],"thumbnail":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1-150x150.png",150,150,true],"medium":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1-300x166.png",300,166,true],"large":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false],"1536x1536":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false],"2048x2048":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false],"woodmart_shop_catalog_x2":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1-600x467.png",600,467,true],"woocommerce_thumbnail":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1-300x300.png",300,300,true],"woocommerce_single":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1-600x332.png",600,332,true],"woocommerce_gallery_thumbnail":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1-150x83.png",150,83,true],"rt_custom":["https:\/\/www.dkmeco.com\/en\/wp-content\/uploads\/2025\/07\/3-1.png",845,467,false]},"rttpg_author":{"display_name":"dkm-admin","author_link":"https:\/\/www.dkmeco.com\/en\/author\/dkm-admin\/"},"rttpg_comment":0,"rttpg_category":"Tableau<\/a>","rttpg_excerpt":"In today\u2019s data-driven world, uncovering insights is no longer enough\u2014what truly matters is transforming those insights into actionable outcomes. Whether","_links":{"self":[{"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/posts\/10318","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/users\/92"}],"replies":[{"embeddable":true,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/comments?post=10318"}],"version-history":[{"count":2,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/posts\/10318\/revisions"}],"predecessor-version":[{"id":10321,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/posts\/10318\/revisions\/10321"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/media\/10319"}],"wp:attachment":[{"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/media?parent=10318"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/categories?post=10318"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.dkmeco.com\/en\/wp-json\/wp\/v2\/tags?post=10318"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}![](\"http:\/\/dkm-website.oss-cn-shenzhen.aliyuncs.com\/upload\/0\/dataBlog\/%E4%BC%98%E5%88%86%E4%BA%AB\/Tableau\/2025.6.20-Tableau%E7%A0%94%E7%A9%B6\/Tableau_Research_jupybara_blog.png\")
\nAs shown above, this is the Jupybara interface. Through its Jupyter Notebook extension format, it supports actionable EDA and data storytelling:<\/p>\n![](\"http:\/\/dkm-website.oss-cn-shenzhen.aliyuncs.com\/upload\/0\/dataBlog\/%E4%BC%98%E5%88%86%E4%BA%AB\/Tableau\/2025.6.20-Tableau%E7%A0%94%E7%A9%B6\/2-%E5%A4%9A%E6%99%BA%E8%83%BD%E4%BD%93%E5%8D%8F%E4%BD%9C%E4%B8%8E%E4%B8%89%E7%BB%B4%E8%AE%BE%E8%AE%A1%E6%A1%86%E6%9E%B6.jpg\")
\nIllustration: Jupybara\u2019s multi-agent collaboration process in data storytelling, showing how the three-dimensional design framework (semantic, rhetorical, practical) is applied in generating and refining analytical narratives.<\/em><\/p>\n![](\"http:\/\/dkm-website.oss-cn-shenzhen.aliyuncs.com\/upload\/0\/dataBlog\/%E4%BC%98%E5%88%86%E4%BA%AB\/Tableau\/2025.6.20-Tableau%E7%A0%94%E7%A9%B6\/%E8%87%AA%E5%8A%A8%E7%94%9F%E6%88%90%E7%9A%84%E5%88%86%E6%9E%90%E6%8A%A5%E5%91%8A%E4%B8%8E%E8%A1%8C%E5%8A%A8%E5%BB%BA%E8%AE%AE.jpg\")
\nIllustration: Jupybara interface showing an automatically generated analysis report and actionable recommendations with highlights for key points and improvement areas.<\/em><\/p>\n