I began by interviewing students and recent graduates to understand their biggest frustrations with interview preparation. The main insights were clear: feedback from AI tools often felt vague or robotic, users wanted to know why they received certain ratings, and the overall experience needed to feel calm and confidence-building, not like a test.

Using these findings, I mapped out conversational flows and micro-interactions in Figma that would mimic a natural coaching experience. From there, I built the AI feedback engine using structured JSON schemas, ensuring each response was consistent, interpretable, and emotionally balanced.

Each AI reply was designed to include:

• A quote from the user’s own answer.

• A 10-point rubric that explained the score.

• A next-step suggestion written in supportive, approachable microcopy.

Early prototypes revealed several challenges. The biggest was AI reliability — responses sometimes lacked structure or emotional consistency. I solved this by refining prompt tone, enforcing schema validation, and adding evidence-based reasoning to every feedback block.

Another challenge was maintaining user trust during latency. Slow responses created uncertainty, so I introduced progress indicators, subtle audio cues, and friendly loading copy such as:

“Give me a moment, I’m reviewing your response carefully.”

These small touches reassured users that the system was thinking, not frozen.

I also tackled error resilience by adding try–catch–finally logic, fallback tips, and clear recovery messages to prevent UI stalls during network interruptions.

Finally, I focused on visual cohesion, blending a technical AI workflow with a calming, confidence-focused interface. Using Raleway typography, soft purple gradients, and rounded components, the interface visually communicates approachability and growth over performance.