Enterprise Knowledge Search · Real-time Analytics
Zelma needed to make fragmented K-12 assessment data usable for parents, educators, and policymakers because the information was scattered across portals and formats and was too hard for most people to query or visualize.
Reported impact
Zelma made school-performance data far easier to query, visualize, and discuss, giving leaders and families faster access to insights that previously took hours or days to assemble.
Zelma turns public education data into a conversational analytics experience by combining a cleaned national dataset with GPT-4 and the OpenAI API. Brown University researchers first gathered and normalized standardized-test data into one format, then Novy used function calling to tell GPT-4 which charts, fields, and views to use when answering user questions. For the Ask a Question interface, Zelma fine-tuned a model to provide data-aware typeahead suggestions and embedded known-good example graphs in a vector database to improve accuracy on edge cases. Questions in plain language enter the interface, the OpenAI-backed orchestration layer maps them to relevant datasets and visuals, and Zelma returns charts, SQL logic, and plain-language context about caveats such as assessment changes. That design makes school-performance analysis faster, more transparent, and more accessible to nontechnical audiences.
Standardized-test data is cleaned into a uniform dataset, user questions are sent through the OpenAI API to GPT-4 with function-calling guidance, and the resulting answers return charts, supporting logic, and contextual explanations grounded in Zelma's education data.
Novy leveraged function calling to tell GPT-4 which visuals and fields to choose from in displaying the data
They embedded and stored known-good example graphs in a vector database to improve accuracy on difficult edge cases
They fine-tuned a model to create a data-aware type-ahead
Zelma offers K-12 superintendents an intuitive, conversational platform to instantly visualize student learning data
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
Related
OpenAI Sample Sep 22, 2025 SchoolAI SchoolAI needed to help teachers personalize instruction and spot struggling students earlier without letting AI simply hand answers to students or create unsafe, opaque classroom behavior. 1 million classrooms reached across 80+ countries Dot · GPT-4.1 · GPT-4o OpenAI Sample Aug 26, 2024 Arizona State University Today, ASU is enhancing educational outcomes by integrating ChatGPT Edu into projects across teaching, research, and operations. Guided by the ASU charter, which prioritizes inclusion over exclusion, research benefiting… ChatGPT Edu 
Lenovo Sample Oct 13, 2026 Formula E Formula E wanted fan experiences that feel immediate, immersive, personalized, and participatory, but its broadcast and event experience was limited by inconsistent onboard video quality, information spread across… Developed in under 30 days. AI Fan Experience Guide · AI Video Optimization · AR broadcast graphics 