After receiving the topic—How can we design an AI agent with a smooth, context-aware experience that helps users learn and handle complex research tasks?—we began by conducting a competitive analysis and accumulated background and context of the project and product. Next, we wanted to understand the real problems users face, so we ran user interviews to identify the key opportunities and entry points for our design.

Competition analysis:

From our competitive analysis, we found that although NotebookLM excels in citation and source retrieval, it is limited in guiding users through structured, project-oriented workflows. Notion and Obsidian, on the other hand, offer strong control over databases and content structure, but their linear editing environments—or the high cost of maintaining non-linear structures—often restrict intuitive thinking.
‍
In contrast, Heptabase provides a unique visual knowledge map that supports spatial thinking and intuitive organization. This revealed a clear opportunity in the market:
Most tools focus primarily on knowledge storage, while Heptabase’s true strength lies in helping users build thinking structures. We believe that combining NotebookLM’s strengths in citation and source retrieval with Heptabase’s visualized knowledge maps would enable AI to act as a thinking partner—one that can surface connections, remove mental bottlenecks, and proactively help users construct meaningful structures. This represents a competitive advantage that is difficult for others to replicate.

Team Collaboration Trade-off
I let go of “everyone must do the full UX process” → shifted to “assign tasks based on strengths.”
Trade-off
Ideally, all designers join research and UI work, but this wasn’t realistic for junior and UI-focused teammates.
Key Decisions
✔ I led UX strategy, research framework, interviews, and decision-making.Junior designers handled briefs and summaries.UI designers focused on high-quality visuals.They joined interviews as notetakers to learn without slowing progress.
Why
1. Maintain stable quality within a tight two-month schedule.
2. Reduce IA/flow decision risks.
3. Enable juniors to contribute where they are strongest and learn effectively.

Research Objective: Understand how users interact with information when handling complex tasks.

User Interview

I was responsible for creating the recruitment survey and developing the user interview questions. I also served as the lead interviewer for all six user interviews.
‍
After sending out the recruitment survey, we received 45 sign-ups within the first week and a total of 67 by the second week. From the responses, I identified potential participants who were likely to engage in complex tasks—selecting three Heptabase users and three non-users. I then conducted screening calls with each person to confirm their suitability for the study. Before the official interviews, I also ran a pilot session to test and refine the interview questions. Based on the pilot and the actual interviews, I continually adjusted the interview guide to ensure clarity and depth.

My phone-screening criteria were based on the research objective:I excluded industry professionals (designers, researchers, or people frequently interviewed for work) and prioritized participants who regularly handle complex tasks and are willing and able to share their thinking.

• HMW help users uncover structure and blind spots in their thinking while organizing complex information—without disrupting their natural workflow?

After that, we narrowed down the ideas through voting. To ensure our resources were invested in the right direction—and to validate whether users actually liked these ideas—we recruited five Heptabase users to test the proof of concept.

I recruited five Heptabase users to participate in concept validation sessions, ensuring our ideas aligned with real user needs. I worked with two UI designers to propose 5–10 ideas to users and create wireframes that visually communicated our concepts based on our brainstorming sessions.

Qualitative Feedback from Concept Validation

😊 Positive Feedback

Users saw clear value in AI-assisted organization:

• AI helps efficiently organize and structure large amounts of whiteboard content.
• Automatically provides structure, categories, and connections users can directly use.
• Especially helpful for research, planning, and case analysis.
• Multi-version suggestions and keyword extraction were considered useful.
• AI Tag speeds up categorization and saves time.

Summary Insight :

AI support at the right moment significantly improves whiteboard organization, structural thinking, and research/planning efficiency.

🤔 Concerns & Frictions

Users also raised doubts about trust and control:

• AI’s categories/structure may not align with user thinking.
• Misidentified key points may increase revision costs.
• Uncertainty about whether AI truly understands relationships between cards.
• Some disliked the Section display format.
• Concerns about data sources and Tag accuracy.
• AI suggestions may sometimes interrupt their flow of thought.

Summary Insight :

If AI misinterprets user intent, it may generate irrelevant structures, create extra work, and disrupt the user’s cognitive flow.

Concept 1 : AI Reviewer

I was responsible for design the UIUX for the AI Reviewer.

The “AI Reviewer” stays quietly by the user’s side when needed.Rather than giving direct answers, it offers different perspectives from multiple AI reviewers to help users break through mental blocks and think more clearly.
It behaves like a coach who can understand the topic you’re organizing on the whiteboard, provide precise suggestions with verifiable references, point out gaps or contradictions, and guide you forward in your thinking process.

Concept 2 : Auto Structure

I was responsible for design the UX for the Auto Structure.

When the whiteboard becomes crowded with cards, connections, and screenshots, the AI Context Architect reads the existing content and proposes possible structures and relationships. It quickly transforms scattered information into a clear, organized hierarchy.
Each suggestion comes with sources and reasoning, and users can choose to accept or ignore them — maintaining full control at all times.