Back to work

Moha Intel

The first build failed in testing. The rebuild became a research workspace analysts trust.

Overview01/

The first version failed in testing. Users couldn't finish basic tasks. The rebuild turned an AI demo into a working research tool for analysts at three companies.

Role

Lead product designer. I owned what the product is. I owned the information architecture and the interaction patterns. I held them through every round of testing.

Team

Lead designer, researcher, frontend developer, backend developer

The challenge

AI chat treats research like a conversation. Research is continuous . Sessions reset, notes vanished, sources were untraceable, so analysts could never build on their own work.

The approach

Rebuild around four persistent surfaces . Show proof at every step. That means visible sources, save states, and version history.

Process02/
01/

The problem

Research teams work in fragments. Tools don't share memory, and AI chat made it worse: sessions reset, notes disappeared, sources untraceable. Research is not conversation. Research is continuous.

D
Drive
No file context
E
Excel
Lost sources
W
WhatsApp
Chat buried
P
PowerPoint
Manual copy-paste
AI
AI Chat
No memory
N
Notes
No traceability
Diagnosis
Severity...
6Tools
0Connections
40%Time lost
Issues found0/6
02/

The first build failed

Users couldn't complete basic tasks. The words confused them. Nothing felt saved. The real problem: researchers needed memory, not a smarter chat.

Version 1.0
Task completion failedcritical
Terminology confused usershigh
No persistencecritical
No session memoryhigh

Researchersneededmemory.

Image 1 of 2
03/

From chat to research workspace

The rebuild swapped chat for four persistent surfaces. Quick Notes for fast capture. Workstation for writing. Channels for topics. Summaries with traceable sources.

Chat tool
01
Fast capture, always saved
02
Primary surface, everything stays visible
03
Organized threads by topic
04
Condensed insights with traceable sources
Image 1 of 4
04/

Validation

Tasks that failed in round one became routine. Users saved freely, built on earlier work, and cited sources with confidence.

Round 2 testing

Validation results

Save without thinkingWork disappeared between sessions
Find any sourceSources were untraceable
Build on yesterdayEach session started from zero
Trust the outputCouldn't distinguish human from AI
0/4 passed
Image 1 of 3
05/

Designing for trust

Trust came from visible proof. Clickable sources, save marks, and version history. Plus a clear line between human input and AI output.

Trust architecture

You

Key risks of investing in Kumamoto?

AI

Population decline acceleratingTSMC inflates land values

Image 1 of 2
Reflections02/
  • Challenges1/3

    The hard problems were words, not technology. One wrong term derailed a whole task flow, and multi-topic research had no existing patterns to borrow.

  • Insights2/3

    The fix was never better AI. It was better feedback: an interface that shows the system's reasoning.

  • What’s next3/3

    The next step: shared channels, live co-editing, and smarter memory. One shared surface.

Swipe to explore

Moha Intel — AI research tool case study