An AI Copilot for human-AI scientific inquiry: Research Architect

Our team is pleased to release Research Architect, an AI copilot for human-AI research collaboration, specifically for hypothesis framing. You are warmly invited to try it at researcharchitect.ai.

When people and AI agents conduct research together, the hardest part is often not performing the analysis, but thinking clearly before the analysis begins. How should the research question be defined? How can existing literature be used systematically? How can expert knowledge be incorporated? How can the main mechanisms be identified? How should the research scope be determined? How can we avoid losing sight of the forest for the trees? And how can we determine what the available data actually cover?

Research Architect helps AI agents avoid proposing hypotheses blindly or becoming overly confident. It also helps them avoid abandoning valuable research directions too early because of incomplete information.

The product uses an influence graph as its central shared medium. AI first identifies and visualizes the factors and relationships relevant to a research question, then assesses the research scope and data coverage. Users can inspect these factors, conduct literature searches, and revise the influence graph using their domain expertise. Based on the improved graph, Research Architect proposes more grounded and testable hypotheses while showing the mechanisms, evidence, and remaining uncertainties associated with each hypothesis.

Since our team introduced the concept of Autonomous GIS, we have believed that AI agents are becoming important executors of scientific research. Human experts serve as a vital interface between the digital and real worlds. They need better tools to identify meaningful research directions, correct the understanding of AI, and allocate research resources more effectively. Research Architect was designed and implemented in response to this need.

The real challenge in building this tool was not software development. It was reviewing our previous projects, studying the research practices of our colleagues, and turning an experience dependent process of research conception into a general workflow that can be expressed, inspected, and revised. To support this goal, we developed a tri-projection framework that moves from the real world to an influence graph, then to research scope, and finally to data. We also distinguish influence relationships as necessary, driving, or modifying.

People familiar with systems science and dialectics may recognize some familiar ideas in Research Architect. We also welcome researchers to share their experience and guide AI reasoning in the same way they would mentor a graduate student.

We will soon share several AI for Science studies and continue exploring how to build new research infrastructure for AI agent swarms. We welcome you to try the platform, exchange ideas, collaborate with us, or discuss investment opportunities.

The tri-projection connects the real world to an influence graph, a defined research scope, and available data

The Research Architect interface: graph view in the center, operation panel on the left, chat box at the bottom, and result panel on the right

Data coverage projected onto the mosquito influence network

Scope assessment: each factor is in scope, partial, or out of scope for the study period and area

A literature review attached to the link from water-holding duration to larval productivity

Generated hypotheses with their confidence, the measurement needed to unblock each one, and a study design

A foundation model gives a general understanding of the world, while the real world is specific and detailed; AI needs context knowledge to serve the real world

More about this project: Research Architect.

First published on LinkedIn on August 10, 2026.