Sachetana — A mental wellness app

Students wanted a place to reflect, but every app they'd tried treated their mood as a data product. The root cause wasn't a missing feature — it was distrust. Once you assume the app is selling your worst days, no amount of gentle copy earns the check-in.

The user story I kept coming back to: "As a student having a rough week, I want to get my thoughts out without wondering who's reading them, so I actually open the app instead of bottling it up."

Before: journaling meant either a clinical intake form or an app that uploaded every word to reflect it back with ads attached. Both raised the cost of a single honest sentence.

After: a two-tap voice or text check-in that transcribes on-device and stays there. The student, not the server, decides what — if anything — is ever shared.

Input: a short spoken or typed check-in, transcribed on-device so raw audio never leaves the phone. Output: a gentle reflection that mirrors what the student said back to them and surfaces one soft prompt — never a diagnosis, never advice.

Human in the loop is the whole point. The model drafts; the student reviews and approves before anything is saved or shared. Counsellors only ever see what a student explicitly hands them.

I scored model responses against a small rubric: does it reflect without diagnosing, stay warm without being saccharine, and avoid any clinical claim? Each check-in sample got a pass/fail on those three, reviewed alongside a campus counsellor.

Responses that drifted into advice-giving — even helpful-sounding advice — failed the rubric outright. Reflecting is the job; prescribing is out of scope by design.

The dangerous failure mode is a wellness app improvising around a crisis. So crisis language never touches the model: it's caught on-device and routed straight to real helpline resources, with no AI in the path.

Everything the model does see is redacted on-device first. If transcription or the model call fails, the check-in still saves as plain text — the student never loses what they wrote.

For a campus, the value is earlier signal at lower cost: students self-reflect before a problem reaches the counselling office, and counsellors spend intake time on the students who opted to share rather than on cold forms.

Trust is the moat. An app students actually keep open is worth more than one with more features they refuse to touch.