Defining the Problem

From the research, I translated user frustrations into clear design goals: manage cognitive load, improve trust, and give emotional clarity. These became five UX goals (G1–G5), supported by design principles that guided how the system should respond in practice for example, explaining reroutes clearly, reducing overload, and matching tone to stress.

From these principles, I mapped seven potential features (F1–F7). Using a priority matrix, I focused on four with the highest emotional and cognitive impact: reroute justifications (F7), filtering irrelevant info (F2), tone-matched reassurance (F5), and delay pacing (F1).

These features became the foundation for the adaptive prototype.

In the ideation phase

I translated defined goals into design logic. Personas grounded decisions in real frustrations, storyboards and task flows then visualised how the adaptive prototype would differ from a static UI in real scenarios. While the Emotion–Logic Flow Map (ELFM) defines when and what the system should adapt based on user state and Screen–State Architecture (SSA) defines when and where those adaptations appear on screen. Alongside these, I applied IBM’s AI 4-Pillars framework (Intent, Knowledge, Reasoning, Data) to shape the AI assistant’s personality ensuring responses were not only functional but also emotionally intelligent and trustworthy.

Prototyping

I iterated from sketches to low-fidelity wireframes, testing assumptions early. User feedback guided the refinement of key features: reroute justifications, tone-matched reassurance, and adaptive pacing. The final high-fidelity prototype captured these behaviours dynamically. The adaptive app allows users to coordinate meetups, find midpoint venues, confirm group decisions in-chat, and even locate friends via AR overlays. AI supports users with emotional tone, pacing, and clarity throughout.

Testing

GPower estimated 34 participants, and 25 completed the study. To evaluate impact, I used a within-subjects controlled study where each participant tested both the static and adaptive prototypes, followed by surveys capturing workload, trust, clarity, emotion, and adaptivity. A mixed-method approach combined quantitative measures (NASA-TLX , Trust Scale, Decision Clarity, Affective Slider) with qualitative insights from open-ended responses.

This triangulation revealed not just statistical outcomes, but also how users described their experience in their own words.

Evaluation

Data was analysed in SPSS using paired-samples t-tests, Wilcoxon signed-rank tests, and thematic coding of open responses.Testing showed that Prototype B outperformed the static benchmark by reducing cognitive load, improving decision clarity and emotional comfort, and being strongly preferred by users.