General conferencing conventions were familiar, but live research needed Discussion Guide, Notes, Key Moments, stimuli, collaboration, role-aware controls and legible AI behavior—without burying the participant.
Back to portfolioResearch technology · Real-time collaboration · Human + AI
Designing a Research-First Conference Tool
A zero-to-one system for conducting human- and AI-moderated research sessions across web and mobile—without making the interface more important than the participant.
Axior Conference Tool is an integrated live-research capability inside the Axior ecosystem. It brings conferencing, research activities, role-based controls and AI-moderated interviews into one session environment.
RoleLead Product Designer
Timeline~2 months
TeamProduct, research, design, engineering, QA
Scope0→1 + two major iterations
StatusLaunch-ready at recorded handoff
Platform5 primary platforms · web + mobile
80+Screens and states designed · approximate
5Primary platforms
~10Role/persona variations referenced
~100UAT issues reported as addressed · approximate
Scope and delivery indicators — not impact metrics.
Executive summary
How I led the design of a research-first conferencing system
The familiar meeting shell was only the substrate. The product had to make research activities, roles and AI behavior first-class without allowing them to overwhelm the conversation.
I led experience direction end-to-end, designed the critical workflows, guided the wider team's output, documented behavior, partnered with engineering, reviewed the build and ran design-led UAT.
Keep the familiar conferencing substrate, then add a research operating layer: contextual capability, explicit role/state logic, responsive prioritisation and visible AI turn-taking.
A zero-to-one five-platform research-conferencing system delivered through two major iterations and implementation/UAT refinement. Post-launch user and business impact remained to be measured.
Context · Ecosystem
The meeting was only one layer of the research system.
Sessions belong to studies. The Conference Tool receives study context, schedule, identity, role and permissions from the surrounding Axior portals and apps.
Axior — Axis of Research
Client Portal · Web
Study contextScheduleRole
Audience Portal · Web
IdentitySessionPermissions
Expert Portal · Web
IdentitySessionPermissions
Audience App · Mobile
ReadinessSessionRecovery
Expert App · Mobile
ReadinessSessionRecovery
Conference Tool
ReadinessWaitingLive researchEnd / rejoinTakeaway
The Conference Tool is an integrated session layer inside a broader research ecosystem, not a standalone destination. Study context arrives with the user; the session hands evidence back.
Problem framing
A familiar call interface could host the conversation—but not the research workflow around it.
The design problem sat in the distance between mature conferencing conventions and research work that was still being assembled across multiple tools.
Mature, expected, cheap to learn.
Audio / videoParticipantsScreen shareChatRecordingCaptions / transcriptHost controlsMeeting states
Deviating here costs trust.
Unowned by the meeting.
Discussion GuideNotesKey MomentsStimuliPollsWhiteboardBreakout activitiesBackroom ChatRole-aware controlsAI moderationResearch context
Usually assembled by hand across three or four tools.
Research tools competed directly with participant visibility for the same canvas.
Moderator, Client/Host and Participant needed genuinely different capability.
Speaking, listening, processing and turn-taking became visible interaction states.
Five platforms multiplied state, responsive behavior and implementation complexity.
Discovery · UX research
We researched where the conversation breaks when the interface has to carry research work too.
Discovery ran on a tight timeline, so it leaned on published evidence, competitive reference analysis, existing workflow analysis and internal expertise rather than a new field study.
Methodology boundary
External-user usability testing had not been completed at this stage. No interview counts, session counts, participant quotes, observed hours or survey responses are claimed.
How the evidence was assembled
Published evidence on meeting friction, AI adoption in research and accessibility failure patterns.
Microsoft Teams, Zoom and Google Meet as conventions users would arrive with.
How research work happens around a live conversation across separate tools.
Product Director, Product Managers and internal Directors on business and research requirements.
Research-domain needs, session mechanics and moderation constraints from the UX Researcher.
Web/mobile feasibility, real-time state, responsive limits and implementation cost.
Stakeholder review, design critique, implementation review and design-led UAT.
Microsoft Work Trend Index, ‘Will AI Fix Work?’ · global workforce survey · External secondary research — not Axior product data.
Live research tooling should reduce memory burden, context recovery and ambiguity—exactly what Guide, Notes, Key Moments, transcript and explicit session state are for.
95%
AI is already ordinary in research practice.
Researchers reported using AI tools regularly or experimenting with them.
3,000+ researchers17 countries
Qualtrics, “The 4 Market Research Trends Shaping 2026” · External secondary research — not Axior product data.
AI moderation is not a novelty story. It raises interaction questions about control, state, trust, pacing, research integrity and recovery.
Research AI is moving inside the workflow.
Qualtrics 2026 Market Research Trends · Current market context · External secondary research — not Axior product data.
Design implication: research AI benefits from living inside the workflow it serves rather than beside it as a detached chatbot.
WebAIM Million 2026 · automated analysis of the top 1,000,000 home pages · External secondary research — not a conferencing benchmark and not Axior product data.
95.9% of home pages had automatically detectable WCAG 2 failures. Accessibility has to be carried by design, component behavior, engineering documentation, implementation review and UAT.
Accessibility is interaction infrastructure.
WHO reports 1.5B people live with some degree of hearing loss and ~430M require hearing rehabilitation services. Captions, transcripts and multimodal system feedback are not optional enhancements in a product whose core unit is a spoken conversation.
Competitive analysis
General meeting tools solved the call. Research teams still needed an operating layer around the conversation.
Teams, Zoom and Google Meet were the historical reference set. Discuss, dscout, Lookback and Great Question are clearly labelled as a current 2026 market-context scan.
Historical project references
Microsoft Teams
What it establishedFamiliar meeting shell, presenter control, whiteboard, collaboration and enterprise governance.
What it left to usFamiliarity reduces learning cost, but research still needs Guide, Notes, Key Moments, stimuli, observation behavior and research-specific AI.
Zoom
What it establishedMature host/co-host models, breakouts, polls, transcript and meeting info.
What it left to usReuse familiar real-time patterns; do not force research workflow outside the meeting.
Google Meet
What it establishedCo-hosts, polls/Q&A, recording, transcripts, note-taking, captions and moderation controls.
What it left to usThe general-conferencing baseline keeps rising. Differentiation belongs in research workflow integration.
Current market context / retrospective scanNot the complete original historical competitor set
Discuss
Human-led and AI-led research in one platform.
Signal: human + AI research in one workflow is an established direction.dscout
Moderated interviews with hidden observers, stimuli, notes and transcripts.
Signal: observation, stimuli, notes and transcript are research workflow, not meeting add-ons.Lookback
Invisible observation room, private back-channel, live streaming, recording and transcription.
Signal: protecting moderator attention while stakeholders observe is a recognised design need.Great Question
Observer rooms, private interview room, recording and repository continuity.
Signal: role separation, observation privacy, status feedback and continuity are active competition.Capability matrix
Where the substrate ends and research capability begins.
Core / nativeResearch-nativeAdjacentNot verified
| Capability | Teams | Zoom | Meet | Discuss | dscout | Lookback | Great Question | Axior |
|---|---|---|---|---|---|---|---|---|
| Core AV conferencing | Core | Core | Core | Core | Core | Core | Core | Core |
| Screen / content sharing | Core | Core | Core | Adjacent | Research | Adjacent | Adjacent | Core |
| Participant / host roles | Core | Core | Core | Adjacent | Research | Research | Research | Core |
| Waiting / admission | Core | Core | Core | Not verified | Not verified | Not verified | Not verified | Core |
| Polling | Core | Core | Core | Not verified | Not verified | Not verified | Not verified | Core |
| Whiteboard / stimuli | Core | Adjacent | Adjacent | Not verified | Research | Not verified | Not verified | Core |
| Breakouts | Core | Core | Core | Not verified | Not verified | Not verified | Not verified | Core |
| Recording / transcript | Core | Core | Core | Research | Research | Research | Research | Core |
| Hidden / private observation | Not central | Not central | Not central | Research | Research | Research | Research | Core |
| Private backroom collaboration | Not central | Not central | Not central | Research | Not verified | Research | Not verified | Core |
| Discussion-guide integration | Not central | Not central | Not central | Research | Not verified | Not verified | Not verified | Core |
| Research notes in live session | Not central | Not central | Adjacent | Not verified | Research | Not verified | Not verified | Core |
| Key moments / highlights | Adjacent | Adjacent | Adjacent | Not verified | Research | Not verified | Not verified | Core |
| Research-specific AI interviewing | Not central | Not central | Not central | Research | Not verified | Not verified | Not verified | Core |
| Explicit AI state / turn agency | Not verified | Not verified | Not verified | Not verified | Not verified | Not verified | Not verified | Core |
| Cross-session research context | Not central | Not central | Not central | Research | Research | Adjacent | Research | Core |
Compiled from official product documentation, 2026, plus verified Axior project capabilities. Competitor rows reflect what public documentation confirms at time of writing, not a feature audit.
Layer 1 · Conferencing substrate
Mic · Camera · Devices · Chat · Participants · Share · Raise hand · Reactions · Recording · Captions · Transcript · Leave/end · Meeting state
Layer 2 · Research operating layer
The strategic gap was not video calling. It was the research workflow around the conversation.Discussion Guide · Notes · AI · Key Moments · Stimuli · Research polls · Multi-document whiteboard · Breakout research · Backroom Chat · Participant management · AI assistance · AI-moderated interviews · Research context · Takeaway questions
Strengths
- Research-first workflow inside the broader Axior ecosystem
- Human and AI moderation in the same product family
- Role/context-aware controls across five platforms
- Research workspace + private collaboration integrated
- Design-system, engineering and UAT ownership through implementation
Weaknesses
- Feature density creates cognitive-load risk
- Role × session × platform × responsive state creates QA complexity
- External-user usability testing was not completed
- Outcome evidence is delivery and qualitative, not post-launch metrics
- Version 2 carried known polish and interaction debt
Opportunities
- 95% of researchers reported using or experimenting with AI
- Movement toward AI embedded in research software
- Research-native competitors validate observation/backroom/notes/AI demand
- Accessibility and transparent AI behavior as differentiators
- Deeper ecosystem continuity from session to research repository
Threats
- General meeting tools keep improving AI and meeting intelligence
- Research-native competitors are expanding interview and repository capability
- AI trust, privacy, bias and turn-taking failures can damage research quality
- Real-time reliability across browsers/devices remains a high bar
- AI patterns change quickly, creating ongoing design debt
Research synthesis
Six patterns changed what the product needed to protect.
Each theme connects evidence → insight → product implication. The product implication is the part that had to survive into the build.
Evidence
Moderators need Guide and Notes while listening; participant visibility must stay primary.
Insight
The interface must absorb complexity rather than ask the moderator to keep managing it.
Design consequence
Participant/content workspace primary; essentials persistent; research tools contextual.
Evidence
Moderator, Host/Client and Participant need different capability; limited roles need contextual access.
Insight
Familiarity should come from a shared shell; safety and relevance from permissions.
Design consequence
One interaction model, role-aware actions, unmistakable public/private separation.
Evidence
Devices, browser permissions, consent, identity and admission happen before the first question.
Insight
Trust can be lost before the conversation starts.
Design consequence
Structured readiness, plain-language recording/consent, explicit waiting/admission and rejoin.
Evidence
Preparing, Speaking, Listening, Waiting, Processing and Paused are required states alongside turn controls.
Insight
A conversational system needs visible state and user-controlled turn boundaries.
Design consequence
Designed state machine, interrupt/hold-floor controls, silence/latency handling and recovery.
Evidence
Guide, Notes, Moments, stimuli, recording, transcript and takeaway belong to one study.
Insight
The session is a stage in a research workflow, not an isolated video call.
Design consequence
Evidence capture inside the session and continuity into post-session review.
Evidence
Five platforms, around ten role variations, responsive states and 80+ screens/states.
Insight
The unit of design is component behavior across conditions, not a screen.
Design consequence
Documented variants, states, permissions and responsive rules.
We thought
We needed an in-house conference interface that could support research features.
We learned
The harder problem was protecting the conversation while research tools, role permissions and AI states changed around it.
The real challenge
Create one familiar live-session model that absorbs research complexity without transferring that complexity to every user.
How might we
Design a research-first conferencing experience that supports different roles, human and AI moderation, and multiple platforms—within a tight delivery timeline?
Behavioral role synthesis
The same live session created four very different jobs.
These fictional composites are built from verified role requirements, session tasks and secondary research. Names, portraits, demographics and quotes are invented for portfolio storytelling; nobody below was interviewed.
Synthetic representative persona · not a research participant
Aditi Mehra
Researcher / Moderator
Senior UX Researcher · Bengaluru · 8 years in UX research · 4–6 remote IDIs a week
“If I'm hunting for controls, I'm not listening.”
Primary job
Run the conversation and capture evidence without losing participant attention.
Highest cognitive riskAttention fragmentation
Needs
- Participant-first hierarchy
- Guide + Notes contextually
- Unambiguous recording state
- Private backroom separation
- Keyboard access
Pain points
- Interface chrome competes with the participant
- Research actions buried in generic menus
- Separate backchannel forces context switching
- Controls that move across devices
Open empathy map
Says
I need the guide visible, but not dominating the interview.
Mark the moment now, analyse it later.
I need to know exactly what the participant can see.
Thinks
Am I still listening deeply enough?
Did the observer see the same moment?
Is recording actually active?
Feels
Responsible for the conversation
Pulled between listening and operating software
Calm when state is explicit
Does
Reads the guide while listening
Captures notes and key moments
Shares stimuli
Reads private prompts
System response
Primary canvas stays dominant
Research tools open contextually
Recording state remains explicit
Backroom is visually distinct
Synthetic representative persona · not a research participant
Priya Nair
General Participant
Operations Manager · Kochi · occasional participant · relies on captions
“Has the host actually seen that I'm here?”
Primary job
Answer naturally without having to learn a research platform.
Highest cognitive riskUncertainty before and between states
Needs
- Plain-language readiness
- Minimal familiar controls
- Explicit waiting state
- Captions
- Clear AI state
- Fast rejoin
Pain points
- Permission prompts feel disconnected
- Waiting silence reads as failure
- Irrelevant controls create anxiety
- AI interruption after a pause
- Connection loss feels final
Open empathy map
Says
Is my microphone actually working?
Who can see or hear this?
Is the AI listening or waiting?
Thinks
Am I already being recorded?
Did the session freeze?
Can I come back if Wi‑Fi drops?
Feels
Cautious before joining
Frustrated by unexplained waiting
Confident when controls are familiar
Does
Checks camera preview
Reads consent
Waits for acknowledgement
Pauses before answering
Rejoins quickly
System response
Readiness checklist + device preview
Host-notified status
Simplified participant controls
Consent before live entry
Rejoin with session context
Synthetic representative persona · not a research participant
Kavya Sharma
Client / Host
Consumer & Product Insights Manager · Mumbai · 11 years commissioning research
“I need to influence the session without becoming part of it.”
Primary job
Keep the study operationally healthy while observing quietly.
Highest cognitive riskRole and channel ambiguity
Needs
- Explicit role treatment
- Admission controls when permitted
- Private backroom separation
- Session health
- Observer-appropriate layout
Pain points
- Observers look like participants in generic meeting tools
- Private talk happens elsewhere
- Permission differences are hard to discover
- Public/private ambiguity is high consequence
Open empathy map
Says
Can I send that probe privately?
Is this visible to the participant?
Has the next participant arrived?
Thinks
Are permissions correct?
Is the moderator aware of my message?
Will the recording be easy to find?
Feels
Responsible for study quality
Cautious around participant-facing actions
Confident when role boundaries are explicit
Does
Monitors waiting and admission
Sends private prompts
Marks moments
Reviews transcript later
System response
Role-aware shell
Distinct private-channel language
Contextual admission
Recording and status continuity
Synthetic representative persona · not a research participant
Dr. Sameer Khan
Expert Participant
Independent strategy consultant · New Delhi · time-constrained · laptop/tablet
“A pause means I'm thinking. It doesn't always mean I'm finished.”
Primary job
Give nuanced expert input quickly without being interrupted or losing context.
Highest cognitive riskBeing interrupted mid-thought
Needs
- Fast readiness
- Visible AI state
- Let me finish talking
- Reliable sharing
- Rejoin continuity
Pain points
- Long onboarding
- AI interruption after natural pauses
- Latency that reads as failure
- Network instability
- Nested menus
Open empathy map
Says
Give me a second, I'm not done.
Is it processing my answer or waiting?
If the connection drops, I need the same interview back.
Thinks
Did it capture the nuance?
Is the AI about to interrupt?
Can I finish from another device?
Feels
Impatient with setup friction
Distrustful when AI behaviour is opaque
Respected when the system waits
Does
Skims instructions
Gives long answers
Pauses mid-thought
Shares a document
Rejoins after network loss
System response
Compact readiness
Visible Speaking / Listening / Processing
Explicit turn controls
Sharing feedback and fallback
Rejoin continuity
Experience lifecycle
Uncertainty started before the conversation—and recovery mattered after it.
Six stages, four actors and the system state between them. No emotional scores: the friction row records what was uncertain, not how anyone felt on a scale.
| Lane | 01 Before | 02 Readiness | 03 Waiting | 04 Live session | 05 Leave / rejoin | 06 Post-session |
|---|---|---|---|---|---|---|
| Participant actions | Receives invite, reads study context in the portal | Grants permissions, previews camera, selects devices, reads consent | Waits, checks status text, may leave | Answers, uses mic/camera/captions, responds to stimuli, holds the floor with AI | Leaves deliberately, or drops and tries to return | Answers takeaway questions when configured |
| Moderator actions | Prepares Discussion Guide and stimuli in the Client Portal | Checks own devices, confirms recording configuration | Reviews who is waiting, decides when to start | Moderates, reads guide, captures notes and moments, shares stimuli, reads backroom prompts | Ends the session or holds it open for a rejoining participant | Uses recording, transcript and moments for synthesis |
| Host / Client actions | Confirms schedule and who is attending | Joins early, confirms observer status | Reviews waiting identity, admits, denies or defers | Observes quietly, sends private prompts, marks moments, watches session health | Steps away and returns without disturbing the session | Retrieves evidence from the study, not a personal notes file |
| System state | Session scheduled, context bound to study | Device and permission state, consent captured, readiness confirmed | “Host notified”; waiting list visible to host only | Live, recording, role permissions, panel state, AI state | Scheduled session still active or ended; rejoin allowed accordingly | Recording, transcript and moments attached to the study |
| Friction / uncertainty | Unfamiliar tool; unclear what the session will involve | Permission anxiety; consent text arriving late or dense | “Did the host see me?” Silence reads as failure | Research tools versus participant visibility; public/private confusion; AI-state ambiguity; screen space | Does leaving end the interview? Can context be recovered? | Evidence scattered outside the study record |
| Design response | Study context carried into the session from the portal | Readiness checklist, preview, plain-language recording notice | Explicit “host notified” status plus visible AV and identity | Three-layer model, role-aware controls, distinct private channel, visible AI state | Rejoin that restores session context while the meeting is active | Takeaway questions and evidence continuity into the study |
Opportunity areas
Five places where design could reduce cognitive cost.
Protect attention
Keep the participant visually primary while research work happens beside the conversation.
← Attention protectionClarify roles and channels
Make what each role can do—and what is public versus private—impossible to misread.
← Role clarityReduce entry uncertainty
Turn readiness and waiting from setup screens into a confident part of the research experience.
← Join confidenceMake AI legible and controllable
Expose conversational state and give the participant authority over turn boundaries.
← AI legibility and agencyKeep evidence in the session
Capture notes, moments, recording and transcript where the research happens.
← Research continuityPrioritisation
What deserved permanent attention—and what could wait until needed?
Placement was argued qualitatively against research value / attention criticality and interaction complexity / implementation cost. No RICE scores existed, so none are shown.
Research value / attention criticality ↑
Persistent · always availableParticipant visibilityMic / cameraRecording stateAI state + turn agency
Contextual · one reveal awayDiscussion GuideNotesKey MomentsScreen shareStimuliBackroom ChatPolls
Quiet defaultsParticipant managementTranscriptWhiteboardBreakout roomsView / settings variants
Interaction complexity / implementation cost →
Retrospective taxonomy / card-grouping reconstruction
Feature cards → task clusters → IA → visibility rules
No participant card-sort study, counts or agreement percentages are claimed.
Join & Readiness
Meeting detailsIdentityCamera previewDevice settingsSpeaker testConsentRecording notice
Core Meeting
MicCameraShareChat
People & Access
ParticipantsWaiting roomParticipant management
Research Workspace
Discussion GuideNotesKey MomentsPollsStimuliWhiteboard
Collaboration & Record
Backroom ChatBreakout RoomsRecordingTranscript
AI Interview & Assistance
AI AssistantAI StateLet me speakLet me finish talkingDone talking
Post-session
Takeaway Questions
How I would validate the taxonomy properly
- Open card sort with representative moderators, clients and participants.
- Hybrid card sort around ambiguous research vocabulary: Stimuli, Key Moments and Backroom.
- Tree testing for findability of contextual tools during a timed task.
- Mental-model comparison between Moderator and Participant.
Information architecture 01
Every session inherits its context from somewhere.
Axior
Client Portal · Web
Study/session context
Audience Portal · Web
Identity · role · permission
Expert Portal · Web
Identity · role · permission
Audience App · Mobile
Session context
Expert App · Mobile
Session context
Conference Tool
ReadinessWaitingLive sessionEnd / rejoinTakeaway
Information architecture 02
The live experience needed an architecture for attention, not just navigation.
Conference Session
Session Context & Status
Study / timingRole / identityRecordingSystem / AI state
Primary Workspace
Participant videoShared content
Essential Meeting Controls
Mic / cameraDevicesShare / captionsLeave / end
People & Access
ParticipantsWaiting / admissionRole actions
Research Workspace
Discussion GuideNotesKey MomentsPollsStimuliWhiteboard
Collaboration & Record
ChatBackroomBreakoutRecordingTranscript
AI Layer
AI Host stateGuide alignmentAI assistanceTurn controlsError / recovery
User flows
Four flows carried most of the product's difficulty.
Failure, denial and recovery paths are part of the flow—not footnotes.
Participant join, readiness, waiting and rejoin
The participant judges the whole study before the first question is asked.
Open session
Study details
Device preview
Permission granted?
Consent + recording
Ready?
Waiting · host notified
Admitted?
Live session
Connection lost?
Rejoin + restore context
Takeaway / End
Host / Client admission
Admission has three outcomes: admit, deny or defer.
Waiting-user notification
Open waiting list
Review identity + context
Admit / Deny / Defer
Update / deny / remain waiting
More users?
Return to session
Moderator research activity
Eight contextual branches, one rule: every action returns focus to the conversation.
Participant primary
Research need
Guide / Notes / Moments / Stimuli / Polls / Whiteboard / Breakout / Backroom
Complete action
Close contextual layer
Participant primary again
AI interview turn-taking and recovery
State must be legible when silence, interruption or latency makes intent ambiguous.
Preparing
Speaking
Let me speak?
Listening
Done talking / hold floor
Processing
Follow-up / next topic
Silence / misunderstood / connection issue
Resume or safe exit
Interaction exploration
Three structures could hold the same features. Only one protected the conversation.
The original exploration wireframes are not included here. These schematics reconstruct the structural argument rather than pretending to be exported historical artifacts.
Rejected
A · Everything persistent
Highest discoverability, but participant becomes the smallest element and every tool competes for attention.
Selected
B · Participant-first + contextual
Participant/shared content stays primary; essentials persist; research tools open on demand over secondary space.
Rejected
C · Research dashboard
Strong operational overview, but turns an interview into software operation and makes participant presence secondary.
Critical design decisions
Four decisions shaped the system. Each accepted a trade-off.
01
One familiar meeting shell. Different capabilities by context.
Seven named roles across five platforms could easily have become several products with different interaction languages.
Evidence
Moderator, Host/Client, Participant and limited roles all need different tools inside the same session, and most arrive with a general meeting tool as their mental model.
Decision
One shared shell; capability computed from role, session type, state and platform.
Trade-off
Consistency comes with hidden contextual complexity: two users in the same session can see different capability.
Design consequence
Same interaction language everywhere; different capability envelope per role. Irrelevant controls are hidden rather than disabled.
02
Research tools could not be allowed to take over the conversation.
Guide, Notes, stimuli, polls and whiteboard all want screen space during exactly the moments when the participant matters most.
Evidence
Research work happens while the moderator listens, not between sessions.
Decision
Three layers: participant/shared content primary; essential controls persistent; Research Workspace contextual.
Trade-off
Some tools are one reveal step away, so discoverability has to be earned through placement and shortcuts.
Design consequence
The participant stays visually primary in every configuration, including the smallest supported screen.
03
AI had to expose conversational state.
In a voice conversation, silence is ambiguous: thinking, listening, latency, failure or interruption.
Evidence
Explicit Speaking, Listening, Processing and turn controls were required product behavior.
Decision
Show the state and give the participant explicit authority over turn boundaries.
Trade-off
Slightly less invisible magic; substantially more legibility and agency.
Design consequence
Let me speak, Let me finish talking and Done talking become research-quality controls.
04
Joining and waiting were part of the research experience.
A participant can lose trust before the first question through permissions, consent, device uncertainty or waiting-room silence.
Evidence
Readiness, waiting, denial and rejoin were all high-consequence transitions.
Decision
Design structured readiness, explicit host-notified waiting, clear denial and context-preserving rejoin.
Trade-off
More deliberate pre-live UI, but less ambiguity and support dependency.
Design consequence
The product communicates what has happened and what happens next before live research begins.
Final experience
One workspace organized around the moment of the research task.
The solution is organized by what the user is trying to do—not by a screen inventory.
Participant · primary workspace1
Discussion GuideTopic 3 · Handover between shifts
Notes & Key MomentsMoment · 12:04Moment · 26:41
2MicCameraShareCaptionsRecordLeave5
Arrive ready
Study context · device readiness · recording notice · consent · join
Wait with certainty
Host notified · waiting state · admission · role and identity
Conduct the conversation
Participant-first canvas · essential controls · recording status · session state
Work while listening
Discussion Guide · Notes · Key Moments · Stimuli · Polls · Whiteboard
Collaborate without leaking context
Backroom Chat · participant management · observer behavior
Recover
Permission · connectivity · sharing · accidental exit · rejoin · role mismatch
Before · fragmented research workflow
Generic meeting tool + four other things.
- Separate Discussion Guide
- Separate notes document
- Separate backchannel
- Post-session evidence reassembly
After · integrated research session
One session, one evidence trail.
- Participant-centered session
- Contextual Research Workspace
- Private collaboration
- Recording, transcript and evidence attached to the study
- Role and AI state explicit
AI experience
AI moderation needed visible conversational state—not invisible intelligence.
Preparing→Speaking→Listening→Processing→Follow-up / next topic
Participant agency
The AI experience was designed and built, not measured. Whether participants correctly interpret each state and whether turn-taking feels natural remain open validation questions.
Validation · UAT
Validation continued after Figma.
External-user usability testing had not been completed at this recorded project stage. Everything below is internal validation: design review, implementation review and design-led UAT—not user research.
Design→Engineering build→Implementation review→SIT / staging→UAT→Fix + retest→Release readiness
≈100UAT issues reported as identified/resolved · approximate delivery-scope indicator
Visual fidelityInteractionsRoles / permissionsResponsive layoutDevices / browsersErrors / recoveryAccessibilityKeyboardWaiting roomMeeting controlsResearch toolsAI statesLeave / end / rejoin
Constraints & trade-offs
Constraints became design criteria, not excuses.
We wantedFull research power in the session
Constraint
Participant visibility is finite screen space
Decision
Three-layer model: primary canvas, persistent essentials, contextual Research Workspace
We wantedCapability tuned per role
Constraint
A consistent interaction model people can learn once
Decision
One shell, computed capability; hide irrelevant controls instead of disabling them
We wantedNatural, invisible AI conversation
Constraint
Silence and latency are ambiguous without visible state
Decision
Explicit states plus three turn controls; less magic, more control
We wantedIdentical experience on every platform
Constraint
Five platforms with genuinely different viewports and input
Decision
Documented responsive priority order rather than pixel parity
We wantedComplete external validation before launch
Constraint
A roughly two-month design timeline
Decision
Internal validation and design-led UAT now; external study plan after launch
We wantedEvery research tool discoverable
Constraint
Feature breadth creates cognitive load
Decision
Progressive disclosure with placement earned by frequency
Collaboration & leadership
Design moved through research, product, engineering and QA—not from Figma to handoff in one jump.
Design Lead — meExperience direction · critical flows
Product & stakeholdersUX ResearchSenior UX DesignerJunior designersWeb engineeringMobile engineeringQA / UATDirectors & leadershipSet the experience direction and three-layer model; designed the critical workflows from scratch; presented trade-offs.
Reviewed and guided senior/junior designer output and maintained cross-role, cross-platform consistency.
Documented behavior, states and permissions; clarified feasibility; reviewed built interactions and corrected mismatches.
Led design-led UAT across web/mobile and iterated after findings through release readiness.
Design system · scale
The interface had to scale across more than screen size.
State × Platform × Role × Permission × Responsive behavior × Accessibility × Error conditions
Meeting controlsParticipant tilesStatus indicatorsLobby notificationPanelsMenusPopoversTooltipsModalsChatDiscussion GuideNotesRecording statesPermission statesError statesResponsive grids
1Active research content
2Participant visibility
3Essential controls
4Role-specific tools
5Secondary actions
Reviewed against WCAG-aligned accessibility requirements: contrast, keyboard navigation, focus order/states, screen-reader labels, captions, touch targets, error/disabled states and system feedback. No formal certification is claimed.
Implementation · handoff
The design stayed accountable through the build.
Design intent
Why the interaction behaves the way it does, not just how it looks.
Behavior documentation
States, variants, permissions, responsive rules and edge cases.
Engineering walkthrough
Web and mobile teams walked the flows before estimating them.
Feasibility clarification
Real-time and responsive constraints changed some designs.
Design review of the build
Implemented interactions reviewed and mismatches logged for correction.
Staging and UAT
Repeated checks until behavior matched documentation.
Outcomes · impact
The delivery evidence was strong. Post-launch behavior was still the next proof point.
Three layers are kept separate on purpose so delivery evidence is not turned into invented product impact.
- 0→1 research Conference Tool
- Human + AI moderation
- Five primary platforms
- Two major design iterations
- Approximately 80+ screens and states
- Around 10 role variations referenced
- ≈100 UAT issues reported as addressed
- Component and behavior documentation
- Launch-ready scope at handoff
- Clearer workflow and role logic
- Meeting-entry friction reduced qualitatively
- Improved consistency across roles and platforms
- Stronger accessibility coverage than the starting point
- Fewer implementation misunderstandings
- Stakeholder approval and major UAT issues resolved
- Task success and time on task
- Participant confidence before joining
- Research-tool discoverability
- AI state comprehension
- AI turn-taking quality
- Mobile / rejoin usability
- Accessibility with representative disabled users
- Adoption and product/business KPIs
System coherence and design accountability.
Experience direction · critical workflow design · redesign after review · design critique · cross-role consistency · component behavior documentation · stakeholder presentation · engineering clarification · implementation review · design-led UAT and iteration.
Production, research context, implementation and validation.
Senior UX Designer · junior designers · product/stakeholders · UX Researcher · web engineering · mobile engineering · QA/UAT.
I did not draw every screen. I was accountable for whether the system held together.
ReflectionStudy 01 · Participant readiness Study 02 · Moderator attention Study 03 · AI comprehension Study 04 · Accessibility Proposed next validation · no results yet
The hardest decision was not which features to include. It was deciding which capability deserved attention in each moment.
What worked
- The three-layer model held under every role and viewport tested internally.
- Computing capability from role/state instead of designing per-role screens.
- Writing behavior early so implementation review was possible.
- Treating readiness and waiting as designed experiences.
What stayed risky
- Feature density remains real; contextual tools can still be missed.
- Stimuli, Key Moments and Backroom vocabulary was never user-tested.
- AI state comprehension remains a hypothesis.
- Version 2 shipped with known polish/interaction debt.
What I would change
- Push for two short external sessions inside the timeline.
- Force a complete permission matrix earlier.
- Test pause-heavy AI turn controls specifically.
- Prototype the mobile session earlier.
What remains unresolved
- Whether one reveal step is cheap enough during a live interview.
- How much AI state is useful before it becomes noise.
- Where Host and Client permissions should diverge.
- Whether the model holds as research features grow.
Proposed next external validation
Device checks, permission comprehension, consent clarity, waiting-state interpretation and forced-drop recovery.
Guide/Notes discoverability, participant visibility under open panels and backroom use without disruption.
Interpretation of Speaking, Listening and Processing; turn-control use; silence and interruption behavior.
Representative disabled users: keyboard paths, screen reader, captions, focus, state feedback and recovery.
The most important design move was not adding more research features. It was deciding when they deserved the user's attention.
Sources, methodology and claim boundaries
Microsoft Work Trend Index — virtual meeting friction, meeting volume and focus time.
Qualtrics 2026 Market Research Trends — 3,000+ researchers across 17 countries; AI adoption and embedded research AI.
WebAIM Million 2026 — automated analysis of the top 1,000,000 home pages; WCAG 2 failure categories.
World Health Organization — hearing-loss prevalence and rehabilitation context.
Historical competitor references — Teams, Zoom and Google Meet.
Current retrospective market scan — Discuss, dscout, Lookback and Great Question.
More work
View all workContinue exploring the system thinking.
Two more case studies from complex product and enterprise design work.
