Inspiration
A person can be exceptionally consistent at work and still struggle to build a healthy routine. The obstacle is not always motivation. Sometimes the health tool itself becomes another job: every metric, every day, or a red score.
The baseline Healthy Living Tracker reflected that problem. It asked everyone to track the same five areas, and incomplete data could look like a poor day. During Build Week, the product was reframed around a simpler question: what is the smallest useful wellness focus this person is ready to sustain right now?
That led to a more focused experience built around user choice, honest data, and AI that assists without taking control.
What it does
Healthy Living Tracker supports five connected facets: sleep, nutrition, exercise, mindfulness, and social connection.
A new user can speak naturally with the onboarding coach. The coach introduces the five facets, listens to what the user wants to improve, and proposes one facet, a subset, or all five. The proposal is not applied automatically. The user reviews it, adds or removes facets, and explicitly confirms the choice.
That choice becomes a product-wide focus contract:
- The dashboard shows only the selected facets.
- Manual and voice check-ins ask only about those areas.
- Scores and trends exclude unselected areas.
- The user can expand or narrow the focus later without deleting history.
For daily logging, the user can speak with the coach or use a manual form. A voice conversation converts the user's words into a structured draft. The AI cannot directly save or overwrite the tracker: the user reviews, corrects, and confirms the draft first.
The app also distinguishes not entered from a genuine zero. It withholds a combined daily score until all currently selected facets have been captured, so an incomplete day is not presented as poor wellbeing.
Voice conversations aim to wrap up around five minutes and stop at ten minutes, with the proposal or draft preserved for review.
How it was built
The app is built with Flutter, Riverpod, and GoRouter. Firebase Authentication handles sign-in, and Cloud Firestore stores owner-scoped profiles and reviewed daily logs. The live voice experience currently uses Gemini Live. A Firebase Function verifies the signed-in user's Firebase token and mints a one-use live credential, avoiding a long-lived AI secret in the distributed app.
The key boundary is model-independent: live AI tools can create only an in-memory proposal or draft. Validation, review, conflict handling, and persistence remain controlled by the mobile app.
The project existed before Build Week as a fixed five-facet tracker with a dashboard, manual entry, trends, and an early Gemini voice check-in. Build Week work added conversational focus onboarding, selected-facet personalization across the product, review-before-save drafts, correct missing-versus-zero semantics, voice time limits, navigation gates, and expanded automated tests.
GPT-5.6 Sol was the model used throughout that Codex workflow. It helped audit the baseline, challenge contradictory product assumptions, trace changes across the data model and UI, implement the cross-layer refactor, diagnose failures, generate and repair tests, and validate the result. The founder made the product, scope, and safety decisions.
The runtime voice path intentionally uses Gemini Live because the product needs a low-latency, full-duplex audio conversation. GPT-5.6 Sol's role was the meaningful product and engineering work performed through Codex, not an in-app voice-model claim.
Challenges
The hardest engineering challenge was making real-time voice reliable on a mobile device. Microphone capture, streamed model audio, WebSocket events, interruptions, app lifecycle changes, deadlines, and unsaved state all interact.
The hardest product challenge was defining what the AI must not do. Prompt instructions alone were insufficient, so restrictions are reinforced in the tool schemas, app-side validation, a separate review step, and automated tests.
Personalization was also cross-cutting. A focus selection changes routing, profiles, forms, AI tools, scoring, charts, empty states, and historical-data behavior. Treating missing data honestly required a data-model change rather than revised copy alone.
Accomplishments
- A complete onboarding-to-insight product loop rather than an isolated AI demo.
- One-to-five-facet personalization across the active experience.
- AI proposals and daily drafts that require user review before persistence.
- Honest missing-data behavior and selected-only scoring.
- A bounded five-minute/ten-minute voice experience.
- A locally verified Android APK.
- Clean static analysis and 44 passing Flutter tests as of July 20, 2026.
What was learned
Personalization is often subtraction, not more configuration.
Natural conversation becomes trustworthy only when paired with explicit transaction boundaries. Propose, review, and confirm proved more useful than letting the model write directly.
Missingness is a product concept, not merely a database detail. A blank, a skipped activity, and a genuine zero communicate different realities.
Codex was most useful as a collaborator across disciplines: it connected a product principle such as missing is not failure to the model, score calculation, interface states, migration logic, and automated tests.
Potential impact
The initial audience is busy, tech-comfortable professionals who want to improve their routines but are unlikely to sustain exhaustive tracking. The impact hypothesis is specific: letting users begin with a manageable focus, speak instead of type, and remain in control of AI-generated data can make repeated self-reflection feel lighter and less judgmental.
This is a wellness product, not a diagnostic or treatment tool. Effects on engagement or health outcomes still need to be validated through user research and longitudinal measurement.
What's next
- Consent-based use of bounded recent history for more relevant reflection.
- Small, editable, goal-linked actions.
- Product analytics for activation, repeated value, and retention.
- Counter-metrics for pressure, anxiety, and unwanted engagement.
- Data export, account deletion, and clearer retention controls.
- Accessibility, local food context, Indian-language support, and user-controlled reminders.
Log in or sign up for Devpost to join the conversation.