Catch the burnout before the crash

Every student I interviewed could describe burnout vividly, but almost none could tell me when it started. That gap is the whole problem. Burnout isn’t a bad day; it’s a slow erosion that the person experiencing it is the last to notice. Andrea described “two phases”, she keeps functioning normally until phase two arrives and the floor drops out. Dani’s crashes “come out of nowhere.” They don’t, the signals were there for weeks; nothing was watching for them.

The cruelest part is what I started calling the discipline mask. UPLB students keep performing while they burn internally. One participant put it flatly: “I don’t care about motivation. I just know I have to submit.” Grades hold steady. Deadlines get met. So every metric a school actually tracks says the student is fine, right up until they aren’t.

I ran contextual inquiries and semi-structured interviews with 12 UPLB students, deliberately skewed toward the people most at risk: first-years still finding their footing, graduate students, working students, and student-athletes and performers. My teammates helped run interviews; I owned the synthesis from there.

I coded every transcript onto an affinity diagram, and the same shape kept surfacing from completely different lives: stress that recovers after a break is normal; stress that doesn’t is the tell. Bianca traced her better semester entirely to actually resting over Christmas. Others described breaks that changed nothing. That single distinction, recovery vs. non-recovery, became the backbone of the whole detection model.

I refused to cluster by year level or program, that tells you nothing about how someone copes. I clustered by behavior: stress response, coping strategy, self-awareness, and openness to digital tools. Two students in the same course can need opposite things. These four archetypes are what the product actually has to serve, and, crucially, they don’t all want to be spoken to the same way.

Nico’s fear, “I don’t want another app that makes me feel like I’m failing”, and Dani’s need for structure sit at opposite ends of one axis. A single tone would either overwhelm Nico or underwhelm Dani. That tension is what forced the central design decision later on.

The whole system turns on one mode prop. Almost every component takes it and flips exactly two things, the accent colour and the corner language, so a screen can never quietly forget which voice it’s in. Amber carries warmth; blue carries data; coral flags risk; green marks recovery. Every suggestion the app surfaces is paired with one of fifteen small illustration motifs by a keyword matcher, so advice arrives as a picture, not another line of text.

The tell isn’t a bad week, it’s a week that doesn’t recover after a break. That’s what separates normal stress from burnout, and it’s what the model watches for.

The personas didn’t just want different features, they wanted to be spoken to differently. So instead of averaging them into something safe and bland, I built one app with two voices, and had it choose silently, from three quiet questions during onboarding, never framed as a personality test. Either voice can be changed later in Settings.

Testers self-sorted by temperament exactly as the personas predicted. If I continued, I’d validate the passive-sensing risk model against a full semester of real data, and pressure-test whether the tone can recalibrate itself without ever telling the user it’s doing so.