ANDY CAO / AI-NATIVE PRODUCT ENGINEER
Ideas intoproducts.
I design, build, and ship AI-native products—from product intent and interaction design to implementation, automated verification, and production learning.
SELECTED WORK / 01—10
Selected Work
Timing method: only development time spent on major product features, integrations, and release hardening is counted. I use active Git days when the feature history is granular, and local development-session timing for repositories imported as a single baseline commit. Maintenance gaps, documentation-only work, and project age are excluded.
01 / IOS PRODUCT / LIVE
Shiguangji
An AI-native food-management iOS app that I designed, built, and shipped from zero to one, turning meal photos and user history into nutrition estimates and personalized guidance.
- Outcome
- Live App Store product
- Core
- Vision LLM · RAG · SwiftData
- Core build time
- Core features + release hardening: 52 active Git days
02 / PRIVATE BUILD / FULL-STACK CMS
PhotographHK
A bilingual portfolio and content system built from zero to one for a photography business. The public site handles project storytelling and inquiries; Payload CMS, PostgreSQL, access boundaries, backups, and automated tests form the pre-launch baseline.
- Core build time
- Frontend, CMS, inquiry flow + delivery baseline: ~20 hours
- System
- Next.js 16 · Payload · PostgreSQL
- Validation
- 24 unit tests · 33 E2E
03 / OPEN SOURCE / FLUTTER
Wujian
A multimodal app for home organization, moving, and inventory. AI turns a photo into a structured draft; the user reviews it before anything is stored locally.
- Core build time
- Major features: 3 active days · hardening: 2 active days
- Reliability
- 21 automated tests and CI
- Principle
- Local-first · Human-in-the-loop
04 / OPENAI BUILD WEEK / WORK + PRODUCTIVITY
SignalForge
Built for OpenAI Build Week, SignalForge turns GitHub signals into explainable SaaS opportunities while separating evidence, model inference, and unvalidated ideas.
- System
- TypeScript · Node.js · SQLite
- Core build time
- Core product + two iteration rounds: ~24 hours
- Boundary
- Explainable scoring and human review
05 / PRIVATE BUILD / AI-NATIVE DELIVERY
Learning Community
Authentication, courses, discussions, and live learning are delivered as verifiable vertical slices. Firebase emulators, Rules tests, component tests, and E2E provide the evidence while people retain product judgment and agents organize, implement, and review.
- Core build time
- Auth, course + live-session slices: 4 active days
- Backend
- Firebase Emulator · Rules · Callable
- Validation
- 31 unit/component · 10 Rules · 4 E2E
constraints→ Specs &
decisions→ Agent
collaboration→ Tests &
review→ Feedback
loop
06 / LOCAL TOOL / FLUTTER
Provenance Lens
A local-first image provenance inspector that parses readable C2PA / JUMBF structures, EXIF, and software markers; it separates what was found from what can be proven and exports a copy without overwriting the original.
- Core build time
- Core local tool: ~30 hours
- Core
- C2PA · JUMBF · EXIF · Fingerprints
- Boundary
- Local processing · no overwrite · careful claims
07 / LIVE SYSTEM / CONTENT OPERATIONS
Personal Publishing System
More than a blog: a bilingual content-production system connecting the public site, photography archive, local CMS, AI-assisted editing and translation, media operations, and quality gates from draft to release.
- Core build time
- Public site, gallery + CMS: 20 active days
- Content
- 24 bilingual pairs · 28 gallery sources
- Quality
- 62 tests · Playwright · Axe · Lighthouse
08 / PRIVATE PROTOTYPE / OPPORTUNITY INTELLIGENCE
ChallengeForge
An opportunity radar that turns public Devpost competitions into actionable candidates ranked by prize density, competition intensity, AI automation fit, and delivery window—while keeping eligibility, IP, and final submission as human gates.
- Core build time
- Opportunity radar + decision pipeline: ~6 hours
- System
- Next.js · D1 · SSE · Public data
- Boundary
- Facts / scoring proxies / human decisions
09 / OPEN SOURCE / DATA + AI
AutoGooglePlay Analyzer
An open-source system connecting Google Play review collection, persistent storage, and LLM Map-Reduce analysis in one pipeline, with a visual Web control surface.
- Core build time
- Core pipeline: 2 active days · dashboard: 1 active day
- Pipeline
- Python · PostgreSQL · LLM
- Outputs
- Markdown · PDF · JSON
10 / RESEARCH PROTOTYPE / HUMAN–ROBOT INTERACTION
Fatigue Recognition
Robot-Assisted Squat Training
Research combining perceived exertion, surface EMG, and platform-recorded kinematics to compare fatigue indicators across experiment design, data collection, and analysis.
- Signals
- RPE · sEMG · Velocity Loss
- Context
- Robot-assisted squat training
- Work
- Experiment, collection, and analysis
