Redefining AI Writing for Global Academia
As the sole product designer at Librum Technologies, I led the end-to-end design of an AI-enabled Intergrity-First Academic Writing System. By pivoting the product strategy toward Academic Integrity and optimizing high-friction workflows, I helped the company broaden its market reach and secure partnerships with international institution.
Design Scope
0→1 end-to-end product design
Status
0-1 Public Launch
Tools
Research & Design Approach
Competitor Analysis, User Experience Design, User Interface Design, Design System, Rapid Usability Testing, A/B Testing, Information Architecture (IA), End-to-end Flow Design, Usability for Complex Systems
Role
Sole Product Designer
My Ownership
The Tension
Education + AI = Trust Gap?
Decision I Drove
No.1 Academic Integrity-first strategy
No.2 Doenba Ecosystem Strategy
Iteration
Information Architecture & Layout
Iteration
Reflection
AI provide guidance and directions, not auto-writing
In-context feedback to reduce comparison effort
Doenba Read was an existing product before I joined. When I started building Doenba Edit, I saw an opportunity to leverage that foundation and create a smoother research-to-writing journey.
Academic work naturally flows end-to-end:
Research (Doenba Read) → Organize sources → Write (Doenba Edit) → Cite → Export
Based on this, I proposed an ecosystem closed-loop model: use citations as the connector so sources captured in Doenba Read are instantly reusable in Doenba Edit during drafting—reducing tool switching and improving end-to-end writing efficiency.
Workflow & AI UX
Iteration
LaTex Compile tool UX
Before
Product framing: Academic Integrity-first positioning
Research synthesis
IA & interaction design
Design & prototyping
Cross-functional handoff
Built and maintained the design system to scale consistency across products
I used rapid usability testing to validate priorities.
Key Findings:
Users relied far more on Library workflows, while AI Chat saw comparatively low engagement.
This led to the iteration direction:
Instead of competing head-to-head with mainstream AI chat tools, strengthen Doenba’s differentiated advantage—high-frequency writing tools embedded in the workspace.
Balancing stakeholder requirment with user needs through a switch tab
In the previous version, the Library area was too constrained, making high-frequency citation work harder than it should be. At the same time, stakeholders wanted to keep AI Chat prominent as part of the product narrative.
After
Inside the panel, Library and AI Chat sit at the same hierarchy level and use a peer feature-switching tabs, so users can quickly switch feature without implying one is “above” the other in the IA.
I organized the AI tools into a workflow that mirrors efficient writing:
Rephrase → Consistency → Review
Key design choices:
Tools appear under the content being revised, enabling 1:1 mapping and reducing comparison effort.
Output is actionable guidance, not rewriting—consistent with Academic Integrity.
If users provide Writing Requirements, suggestions align with those criteria (rubric-like review).
The original compile experience only supported full-document compile in the Export stage, which is too expensive for frequent “quick checks.”
So I move formatting support closer to the writing process so users can see, understand, and recover before final export.
The most valuable AI experience is not always a chat interface
Usability testing showed that users relied on Library and citation-related actions more frequently than AI Chat. This reminded me to prioritize the user’s core workflow over the most visible technology pattern and to use AI where it provides contextual value rather than making it the center of every interaction.
Complex tools should match system cost to the scope of the user's intent
My attention had been on interface feedback whether errors were visible, whether recovery was clear, not on what each check cost to run. Requiring a full document compile for every local check meant a single equation and a full layout review triggered the same wait, the same interrupted focus, the same computing load. The fix was using the smallest sufficient operation for the task at hand, not the most complete one.
In education, many AI writing tools are positioned around rapid output. They framed as shortcuts rather than skill-building. In academic settings, the market often swings between two extremes: “done-for-you” writing tools and policing tools that detect AI use.
Students may use AI to improve learning—or to cheat. Meanwhile, professors and institutions are skeptical due to academic integrity concerns. This trust gap limits adoption of many AI products in academic settings.
I proposed shifting Doenba Edit away from “AI as a shortcut” toward Academic Integrity-first learning support.
The AI should behave like a professor’s 1:1 review: it doesn’t write for the student, but provides actionable guidance while preserving authorship and control.
Design principles derived from this strategy: