Load Away

1. 01

   Built the ARKit gaze pipeline: calibrated face→camera classification, the eye-gaze peek guard, the eyes-closed path, the frozen-mesh guard, and the continuous player identity lock

2. 02

   Built the watchOS companion and the heart rate path: HKWorkoutSession and HKLiveWorkoutBuilder streaming BPM over WatchConnectivity, with a stale-sample watchdog and a resting-low baseline

3. 03

   Diagnosed the watch pairing failure down to build configuration, embedding the watch app as a companion with WKApplication and WKCompanionAppBundleIdentifier

4. 04

   Worked as one of three engineers in a five-person team alongside two designers

1. 01

   Gaze-gated loading bar driven by ARKit face tracking

2. 02

   Calibrated, face→camera relative gaze classification instead of absolute head angle

3. 03

   Eye-gaze peek guard with a peek tax that reverses progress

4. 04

   Eyes-closed counted as looking away, gated by distance and pitch

5. 05

   Continuous player identity lock preventing a face swap mid-round

6. 06

   Level 2 "Unstable Loading": jitter, random drops, checkpoints, and a fake completion at 99%

7. 07

   Roughly two-second window to look back at 100% before the bar collapses

8. 08

   Apple Watch heart rate feeding bar volatility, shipped off by default

9. 09

   Spoken taunts from a static phrase bank with pre-baked voice clips, optionally refined by on-device Foundation Models

10. 10

    Fake push notification trap timed to bait a glance back

11. 11

    English and Indonesian localization across UI, taunts, and voice

12. 12

    Manual "Hold to Look Away" fallback for devices without TrueDepth

A native SwiftUI app with no game engine. Gameplay is a @MainActor GameEngine on a roughly 30 Hz timer, structured as MVVM with a state machine, and every threshold is tunable through ConfigService rather than hardcoded. Gaze comes from ARFaceTrackingConfiguration, heart rate from a watchOS workout session over WatchConnectivity, audio from pre-baked clips with AVSpeechSynthesizer as live fallback.

Front end

Swift, SwiftUI, Swift 6 Concurrency

Tools

Xcode, Figma

Integrations

ARKit, HealthKit, WatchConnectivity, watchOS, Foundation Models, CoreHaptics, AudioToolbox, AVSpeechSynthesizer

1. 01

   Problem

   TrueDepth stops tracking near profile (~45°+) and reports the face as lost, so a full head turn read as "face lost" rather than "looking away".

   Solution

   Set the away-gate at 50°, well inside the reliable range, and enforced no-peeking with an eye-gaze guard instead of a wider angle.

2. 02

   Problem

   Judging head angle in ARKit world axes drifted with phone motion and inverted the reading, turning right went undetected while turning left registered as a peek.

   Solution

   Re-derived the whole classification as face→camera relative geometry, which depends only on where the camera sits in the face's own frame and stays drift-immune in landscape.

3. 03

   Problem

   Players cheated by turning their head away but sliding their eyes back to the screen, which a pure head-angle test cannot catch.

   Solution

   Added a pose-independent eye-gaze peek guard comparing the eye ray against the ray to the camera, plus a peek tax that knocks the bar backwards.

4. 04

   Problem

   A perfectly still face mesh signalled a stale or replayed anchor whose frozen pose was often a turn, which falsely advanced the bar.

   Solution

   Added a frozen-mesh guard that treats a non-jittering mesh as face-lost.

5. 05

   Problem

   ARKit silently re-fits a surviving anchor onto a different face during a player swap, so UUID alone let a substitute inherit the lock.

   Solution

   Added a continuous identity check using a canonical-mesh shape signature and an averaged pre-lock fingerprint, run at every angle the mesh stays reliable.

6. 06

   Problem

   There is no realtime heart rate stream from the phone; live BPM requires a workout session, which lives on the watch.

   Solution

   Moved heart rate onto a watchOS companion running HKWorkoutSession and HKLiveWorkoutBuilder, streaming BPM over WatchConnectivity with a stale-sample watchdog and a resting-low baseline.

7. 07

   Problem

   A watch app that looked installed never paired; WCSession.activationState stayed at 0 and heart rate never reached the phone.

   Solution

   The fix was build configuration, not code: embed the watch app as a companion inside the iOS app with WKApplication and WKCompanionAppBundleIdentifier set, then erase the stale simulator install shadowing the rebuild.

8. 08

   Problem

   The eyeBlink blendShape reads high even with eyes open when the face is far away or tilted, so eyes-closed detection produced false positives.

   Solution

   Gated eyes-closed detection by a maximum distance of 0.6 m and a 20° pitch delta.

1. 01

   The central assumption, that looking away is just an angle, was wrong. Calibration and face-relative geometry were mandatory, not polish.

2. 02

   The hard bugs were physical sensor behaviour and Xcode target wiring, and only surfaced on real hardware. Documentation and AI described the happy path; the device described reality.

3. 03

   Anti-cheat signals doubled as comedy: because the camera already knew the eyes, brow, and heart rate, frustration taunts and the notification trap came nearly for free.

4. 04

   Cutting Level 1 from play sharpened the hook even though it was built and working.