Design AR Virtual Try-On - System Design | WinJob

Problem Statement: Real-Time Garment Overlay

Design AR Virtual Try-On so shoppers anchor eyewear or apparel on a live body mesh with commerce handoff. Interviewers at Warby Parker, Sephora, and Nike expect you to quantify 8M DAU, 350k peak concurrent sessions, and 4.2M tracking events per second before naming microservices.

Why this matters in the understanding phase

Candidates who upload camera video to “improve fit” fail privacy reviews and miss sub-33ms glass-to-glass budgets. try-1-sec-01-phase-hook ties problem statement: real-time garment overlay to concrete APIs and stores.

Focus 1.1: live camera passthrough with world-locked frames — invariant try-1-sec-01-focus-0
Focus 1.2: eyewear and apparel SKUs in one catalog — invariant try-1-sec-01-focus-1
Focus 1.3: shareable snapshot without storing raw video — invariant try-1-sec-01-focus-2
Focus 1.4: commerce handoff to cart within two taps — invariant try-1-sec-01-focus-3

Operational detail

State explicitly: CP for published garment manifests; AP for body estimates and tracking deltas.

Failure and edge cases

low texture background — widen landmark search and pause garment swap (try-1-sec-01-f0)
occluded ears on bulky hair — hide temple arms until confidence recovers (try-1-sec-01-f1)
stale GLB after merchandising update — force manifest version bump (try-1-sec-01-f2)
duplicate POST /sessions from flaky mobile network — idempotent sessionId header (try-1-sec-01-f3)

Interview checkpoints

javaOne Dark Pro

1 public record GarmentManifest(String sku, long version, String glbUri, String rigHash) {}

pythonOne Dark Pro

1 def tracking_confidence(residual_px: float, light_lux: float) -> float:
2     return max(0.0, 1.0 - (residual_px / 12.0) - (max(0, 200 - light_lux) / 400.0))

typescriptOne Dark Pro

1 export function shouldFallback2D(confidence: number, threshold = 0.55): boolean {
2   return confidence < threshold;
3 }

Why interviewers care

AR Virtual Try-On interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Problem Statement: Real-Time Garment Overlay that proves you understand real outages, not happy-path diagrams.

Key Highlights

•live camera passthrough with world-locked frames
•eyewear and apparel SKUs in one catalog
•shareable snapshot without storing raw video
•commerce handoff to cart within two taps

1	def tracking_confidence(residual_px: float, light_lux: float) -> float:
2	return max(0.0, 1.0 - (residual_px / 12.0) - (max(0, 200 - light_lux) / 400.0))

1	export function shouldFallback2D(confidence: number, threshold = 0.55): boolean {
2	return confidence < threshold;
3	}