Inspiration
The most important parts of life are often managed by the least coherent tools. Our calendar knows meetings but not why routines matter. Pet-health details disappear across chats, receipts, and memory. Finance apps show balances but rarely explain risk without demanding sensitive cloud data.
Yuanbao Life OS began with a simple conviction: a personal life platform should be comprehensive without being invasive. The result is a suite of three native Android apps—Yuanbao Schedule, Yuanbao Life, and Yuanbao Finance—that treats time, care, and money as connected parts of one human life while keeping the data local and the logic inspectable.
What it does
Yuanbao Schedule combines day/week/month/four-quadrant planning, recurrence, range dates, alarms, overdue rollover, completion rules, widgets, system calendar integration, focus sessions, and a remarkably capable free-form notebook with drawing, grids, images, floating text, layers, locking, and undo/redo.
Yuanbao Life gives pet and family care a dedicated memory: weight, vaccination/deworming/dental care, meals, footprints, and milestones each get purpose-built inputs and timelines instead of being buried in generic notes.
Yuanbao Finance turns portfolio screenshots into explainable insight. Bundled Chinese ML Kit OCR runs on-device, then a human review screen prevents low-confidence rows from silently entering the portfolio. The app visualizes allocation, diagnoses risk exposure/concentration/diversification/cash buffer, tracks five major index valuation temperatures, applies simulated scheduled contributions on official trading days, and surfaces market/policy news. It never places an order and never pretends a model is financial advice.
All three apps are independently installable, require no account/backend/API key, store primary data in Room, and export human-readable JSON.
How we built it
The suite is a Kotlin/Jetpack Compose multi-application Gradle project targeting Android API 34 and supporting API 26+. Each app has its own package name, database, resources, workers, and APK. State flows from Room-backed repositories through deterministic domain rules and ViewModels into Compose UI. WorkManager handles finance refresh and planned-contribution bookkeeping; AlarmManager, widgets, and CalendarProvider connect schedule data to Android; ML Kit performs device-side OCR.
Codex with GPT-5.6 was the core engineering workflow, not a last-mile assistant. We expressed product behavior conversationally, then Codex inspected the existing design system, researched authoritative constraints, edited across entities/DAOs/repositories/domain rules/ViewModels/Compose/tests/resources, ran Gradle, diagnosed errors, and iterated. This compressed what would normally be a sequence of handoffs—product specification, architecture, implementation, QA, and packaging—into one traceable loop.
Key Codex + GPT-5.6 decisions
- Converted an ambiguous date-picker requirement into a deterministic, regression-tested state machine: first tap selects one day, second tap creates an ordered range, third tap restarts.
- Chose on-device Chinese OCR plus explicit human confirmation rather than uploading sensitive brokerage screenshots.
- Separated “valuation temperature” from licensed historical percentiles and disclosed every model boundary.
- Designed an auditable portfolio score from risk exposure, concentration, diversification, and cash reserve instead of an opaque AI verdict.
- Made recurring planned contributions idempotent with a unique
(plan, tradingDate)ledger and an official year-versioned exchange calendar. - Repeatedly validated unit tests, Lint, Debug/Release compilation, APK metadata, and signing as part of the conversation.
Challenges
The hardest challenge was not writing more code; it was refusing false convenience. Finance features had to be useful without overstating data authority. OCR had to save effort without silently corrupting a portfolio. Recurrence, rollover, range selection, and background work had to remain deterministic across process restarts. The notebook had to combine text, drawing, images, layers, gestures, and persistence without collapsing into a toy canvas.
Accomplishments we are proud of
Yuanbao is a coherent, runnable product rather than an AI-themed mockup. Three independent apps compile together, install side by side, share a recognizable visual language, expose testable deterministic rules, and make privacy visible in the interaction design. The finance module is particularly unusual: it combines offline OCR, explainable analysis, official-calendar scheduling, and explicit model disclosure in an approachable consumer experience.
What we learned
GPT-5.6 and Codex are most powerful when they participate in decisions, validation, and iteration—not just code generation. The best results came from giving Codex real constraints, letting it inspect the entire product, and treating tests and disclosures as first-class features.
What's next
Next we will add optional encrypted device-to-device sync, a shared dashboard that preserves independent installability, English localization and accessibility passes, more import adapters, signed reproducible releases, and future year-versioned trading calendars.
Built With
- android
- codex
- gpt-5.6
- jetpack-compose
- junit
- kotlin
- ml-kit
- room
- workmanager