This is a submission for the Gemma 4 Challenge: Build with Gemma 4
What I Built
I built QuietPulse, a privacy-first, on-device AI journaling application designed to bridge the gap between daily mental health tracking and clinical psychiatry.
Patients suffering from anxiety or depression often experience severe memory bias (e.g., recency bias, exaggerating bad days, or forgetting details) when answering standard clinical questionnaires like the GAD-7 or PHQ-9 at the doctor's office.
QuietPulse solves this by acting as a daily voice journal and a "Questionnaire Assistant." Users can tap a button and dictate their daily thoughts into the app for as long as they want. The app processes these journals locally to extract underlying emotional states and clinical markers. When it's time for their doctor's appointment, the app can aggregate the last days of data into an objective timeline. If the questionnaire asks, "How many days have you felt on edge?", QuietPulse provides the exact number of days the criteria was flagged in their journals, complete with specific AI summaries as evidence, ensuring the psychiatrist gets perfectly accurate data.
I'm taking this competition as the kick off to make my repo public and open source, I will be adding a lot more features, but I want clinics to be able to take this app and provide it for free to their patients and so they enrich the questionaries available for others to use, that way we all benefit, fighting with depression and anxiety is tough and anyone deserves help, this is my way to contribute.
Demo
Code
https://github.com/fhuezo/quiet_pulse
How I Used Gemma 4
Mental health data is arguably the most sensitive data a person can generate. Routing highly intimate, daily audio journals to a cloud API is a massive privacy risk and a non-starter for many users.
I used the Gemma 4 E2B (Edge 2-Billion parameter) model running entirely offline on-device via the MediaPipe LLM Inference API in Flutter. This was the perfect fit for the use case because it guarantees absolute privacy—the user's voice and text never leave their physical hardware.
Despite being small enough to run smoothly on a mobile device, Gemma 4 is incredibly capable at complex natural language understanding. I prompt the model to analyze the unstructured, often rambling voice dictations and output a strict JSON payload containing:
An intensity-mapped array of detected emotions.
A concise 1-sentence clinical summary.
Specific detection flags for GAD-7 (Anxiety) and PHQ-9 (Depression) criteria.
Gemma 4 processes this perfectly on the edge, powering the entire clinical aggregation engine without a single network request.
