Taylor P1, Ross S1, Gao C, Cheng B, Osborne D, Bunn E, Shen R, Harvison E
1RMIT University, Melbourne, Australia
Biography:
Patrick and Shiv come from the two RMIT teams that designed and built ResearchMate:
Abstract:
The RACE (RMIT Advanced Computing Ecosystem) team, whose mission is to provide cloud-oriented services, training, DevOps, and expert support that enable RMIT staff, students, and industry partners to maximise their research impact; and The RMIT Continuous Improvement and Systems team, whose aim is to optimise operational efficiency through process redesign, system upgrades, and data analysis.
This project is the result of a strong collaboration with valuable input from the researcher community who participated in the pilot program. Together, we've built a tool that makes research information more accessible and intuitive for RMIT's research community.
Researchers spend hours each week navigating fragmented institutional knowledge. Answers to questions about ethics, funding, intellectual property (IP), Higher Degree by Research (HDR) milestones, or data management exist, but are buried across policy documents, portals, and guides that were never designed to work together. Multiplied across a large institution, this becomes a structural cost to research productivity. ResearchMate addresses this with a conversational AI assistant embedded in Microsoft Teams, a tool researchers already use. It delivers instant, source-cited answers around the clock, removing the need for additional logins or portal navigation.
The platform combines three layers. An agentic reasoning layer, built on Amazon Web Services (AWS) Bedrock with Anthropic's Claude model, autonomously selects tools at runtime: a curated institutional knowledge base, live web search, or acronym lookup. It chains these to construct responses and flags low-confidence queries for human follow-up. A mandatory trust layer attaches direct source citations to every response, directly addressing the hallucination risk that makes generic AI unsuitable for research environments. An institutional intelligence layer categorises each conversation, surfacing knowledge gaps via a live analytics dashboard and only stores metadata, to preserve user privacy.
In a controlled pilot with approximately 90 RMIT academics, ResearchMate reduced the time spent navigating institutional processes, with users consistently highlighting faster access to trusted, context specific information and increased confidence in the information provided. This presentation shares the architecture, pilot outcomes, and practical lessons for institutions deploying trusted, domain-specific AI to reduce confusion and improve productivity.
