Assistance, with room for judgment.
Two connected interests: using AI to support design work, and designing AI-assisted products that keep people informed and in control.
What these examples are
Eight authored, removable demonstrations prepared for this portfolio review. They are not client case studies, research findings or model transcripts. There is no live inference and no upload. Personal professional examples can replace them after review.
AI during design
An initial suggestion is a starting point. The design work is to define the task, trace the evidence and correct what does not hold.
Clarify the brief
Illustrative demonstrationA synthetic team asks for “a smarter reminder page.” That request does not say who uses it, what the system knows or which decisions must stay with a person. Generating a polished layout at this point would hide the missing brief.
Authored brief revision
Goal: set, inspect and change one appointment reminder. Known: appointment date and chosen contact channel. Constraints: no invented appointment details, no automatic rescheduling, accessible labels. Required states: empty, saved, editing, permission unavailable and recoverable error.
Reject the unsupported assumption that “smart” means automatic. Define what can be assisted and what needs confirmation before discussing layout.
Trace a synthesis
Illustrative demonstrationA synthesis should preserve where a theme came from and where the material disagrees. The notes below are authored illustrations, not participant quotations or findings from a study.
Illustrative notes only
N1: A frequent editor wants the time change close to the reminder summary. N2: A first-time user wants an explanation before changing a repeating rule. N3: A person using a screen reader needs the selected schedule announced.
Proposed theme: make editing easy to locate. Trace: N1 and N2 concern editing, but imply different support needs. N3 concerns accessible state, not frequency of editing.
Do not flatten these notes into “all users want fewer steps.” Keep the contradiction visible and separate accessibility requirements from a preference hypothesis. Actual research would need its own consent, evidence and interpretation.
Compare flow alternatives
Illustrative demonstrationA synthetic reminder editor can show every rule at once or reveal rules after a familiar choice. Neither is automatically better. The comparison needs discoverability, effort and recovery, not only the number of screens.
A · All rules together
Show frequency, interval, weekdays and time in one form. Expert editing may be direct; first-time setup carries more reading and more irrelevant choices.
Recovery requirement: retain all entered values and explain conflicting rules next to the affected controls.
B · Start with a routine
Choose a familiar frequency, reveal the relevant rules and review the result. Initial choices are lighter, but the advanced path must be explicit.
Recovery requirement: go back without losing the schedule and edit from a readable summary.
Choose against the task and available evidence. A shorter path that hides the needed choice can be harder than a longer, understandable one. These alternatives are authored concepts, not measured test results.
Write across languages
Illustrative demonstrationA translation should preserve what happened and what can be done next. It should not add a promise or remove the distinction between saving a setting and delivering a notification.
Authored copy pair
Reminder time saved. You can change it in reminder settings.
Notifications are off. Enable them in device settings, or keep this reminder without notifications.
Erinnerungszeit gespeichert. Du kannst sie in den Erinnerungseinstellungen ändern.
Benachrichtigungen sind ausgeschaltet. Aktiviere sie in den Geräteeinstellungen oder behalte diese Erinnerung ohne Benachrichtigungen.
Replace a generic “Something went wrong” with the condition and a real next step. Check the longer German copy at narrow widths. These examples describe a hypothetical reminder product, not a tested MyTherapy permission flow.
Prototype with care
Illustrative demonstrationThis portfolio is itself a collaboration artifact: a source brief, supplied screenshots and authored bilingual content have been assembled into editable pages. It is not evidence that Oksana used this workflow on a client project.
What matters beyond a layout
A case link must open a case. A language link must preserve the story. A CV state must not promise a missing file. Screen captions must say what the image actually shows. A generated sentence must not turn a draft into a verified achievement.
Separate source delivery from acceptance. Semantic structure, interaction requirements, responsive layout and claim limits belong in the brief. Oksana’s review still needs to establish whether the voice and examples represent her.
Designing AI-assisted interfaces
Make sources visible, expose disagreement and let the person correct the result or take another path.
Compare fields with evidence
Illustrative demonstrationA document-assistance concept, connected to the Gini challenge context in the supplied material. It is not employment, a shipped product or a reconstruction of the challenge screens. This synthetic invoice/transaction example shows why field evidence belongs beside a proposed value.
Concept · Synthetic sources
Synthetic invoice
€84.90
4 October 2026
Synthetic transaction
€89.40
5 October 2026
Inspect the disagreement, edit the fields below and record a local review. No “correct” match is inferred by this prototype.
Do not disguise disagreement with one confidence badge. Compare amount and date separately, preserve the original sources and make the proposed values editable.
Keep a manual way forward
Illustrative demonstrationA system that cannot resolve a mismatch should say so without trapping the person. In this authored scenario, the two sources disagree and the interface does not know whether they belong together.
Concept · Needs review
Amount and date differ. Keep the sources separate until you have checked them. You can continue manually; nothing will be matched automatically.
Open manual review
Synthetic invoice
€84.90
4 October 2026
Synthetic transaction
€89.40
5 October 2026
Use your judgment to enter a reviewed value. Closing this section preserves the fields; Reset restores the teaching example.
Keep actions reversible and name their actual effect. “Record local review” records a demonstration state; it does not approve a payment or teach a model. A production system would need additional domain-specific safeguards.
Evaluate the difficult states
Illustrative demonstrationAn evaluation checklist should include the moments when assistance is least reliable. The checks below are editable local selections, not a report that a system has passed an evaluation.
Review topics, not performance scores
Select the topics you want to include in a review.
Evaluate the whole recovery loop, not just a plausible first answer. A real evaluation needs defined tasks, representative evidence and recorded findings. No accuracy or productivity result is claimed here.
Sources, and their limits.
Microsoft HAX and Google PAIR inform these editorial choices about capability boundaries, evidence, correction and recovery. They do not verify Oksana’s professional AI use or the behavior of an actual model. The examples are authored teaching material.