GenUI decision workspace
Role
HCI Designer
Timeline
3 months
Deliverable
MSc thesis
Tags
GenUI
Deterministic
Prototype
For my MSc in human-computer interaction, I designed and evaluated a generative UI for decisions people come back to for weeks, replacing the tools they improvise from tabs, carts, and messages to themselves. AI reorganises that scatter as intent shifts, with the user in control. The catch: the regeneration that makes it useful also disorients people. So generation stays in predictable zones, and the interface does the maths, recasting lists as tables, matrices and boards. Participants weighed options with less effort and more confidence.
Decisions
D1
Stable by default, regenerate on intent shift
People collect options in a list, then jump to a spreadsheet the moment they switch to comparing. The scannable list stays as the resting state, and the same item set regenerates into a table or matrix as intent shifts. The interface follows the decision as it changes.
The list becomes a dimensions table, then a spatial matrix.
D2
Constrain generation to predictable zones
Predictability is what makes "the UI changes itself" tolerable. Adaptive interfaces that shuffle elements around disorient people. So generation is deterministic and confined to bounded zones: the page, the panel's position, and the criteria layout never move, and only the output zone redraws. The arrangement is all that's ever generated. Every fact in the panel is pulled from its source or computed from it, leaving the interface no way to invent content.
The criteria row stays put; only the zone above redraws.
D3
Liberal input, structured output the user approves
Options arrive as URLs, notes, and screenshots, and the context between them goes missing. Entry is liberal and every output structured: "£1500 for sofa and table" typed in natural language re-renders into chips the user approves before they apply, and each new line stacks more onto the set. Dropped links become item cards in the same list, so intent stays visible and editable on the panel.
Dropped links become item cards; typed text becomes chips to approve.
D4
Reversible 'what-if' with smallest-change suggestions
Criteria are chips that cost nothing to add, drop, or adjust, so exploration stays low-risk. The design effort went on the dead-end: when no combination fits, the interface suggests the smallest change that would work, shown as nested chips, so the user keeps moving.
Raise the £1,500 budget to £1,638, or add a house brand.
D5
Compute the trade-off as conditions change
The interface does the arithmetic that pushes people into spreadsheets. Each pairing keeps a running total that the delivery toggle recomputes on the spot. Set a budget and every combination is totalled against it, so the pairings that fit stay visible as fees or stock shift.
Flip Excl. delivery; a fifth pairing comes into budget.
Results
Participants trusted the interface's working because it stayed visible; they said it clarified the decision and imagined the same tool for gardens, gifts and holiday lets. A scaffold that reinforces the person's reasoning rather than replacing it.