Global social platform reaches 99.6% moderation accuracy on Amazon Bedrock

Well-Architected ReviewAI safety guardrailsModel evaluationModel routingModel fallbackProvider abstractionPrompt engineeringEvent-driven architectureUnit-cost optimizationInfrastructure-as-code foundationAgentic workflowsCost-optimization deep diveHigh-risk-issue remediation roadmapTarget-state architecture design & costed POCDay-2 runbooks & team enablementInfrastructure observabilityPrivate connectivity & network isolationSecrets management & credential eliminationModel cost governanceWorkload placement & runtime selectionStorage tiering & lifecycle policies

54 → 99.58%

NSFW accuracy — prompt engineering alone, no fine-tuning

~$1

per 1,000 images, end to end

4 of 900

NSFW images missed — 99.56% recall, the number that matters when false negatives are unacceptable

Amazon BedrockAmazon EKSAmazon EMRAmazon CloudFront

A global consumer social platform lives or dies on the safety of the media its users upload. VeUP delivered and operates the media-platform backbone of that promise: multimodal content moderation on Amazon Bedrock with a measured evaluation harness, a synchronous image path and an asynchronous event-driven video pipeline, Amazon EKS and Amazon EMR Spark analytics run under managed FinOps, and Amazon CloudFront delivering the approved content worldwide.

The challenge

A user-generated-content platform is a media company whose entire catalog arrives unreviewed, around the clock, in every format users can produce. The platform’s safety policy demands that unsafe content never reach users — an editorial standard enforced at machine speed — while the same estate must serve approved media globally with consumer-app latency and keep the analytics and compute footprint economically sane as volume grows. Moderation, media processing, analytics, and delivery had to work as one system, with measurable quality at the moderation gate.

The solution

VeUP built the moderation gate as a data-science-grade media pipeline. Amazon Bedrock’s Qwen 3 VL multimodal model is the primary inference engine, with Amazon Rekognition preserved behind an engine-router abstraction exposing one verdict contract across both — a synchronous low-latency path for images and an asynchronous event-driven pipeline for video and bulk media, orchestrated by AWS Step Functions and Amazon EventBridge Pipes with per-frame judgments aggregated K-of-N into a single verdict per clip. Engine selection was evidence-based: a head-to-head benchmark over 1,501 labeled images measured recall and unit cost before production promotion. Around the moderation gate, the platform’s Amazon EKS workloads run on Karpenter with Spot capacity and its Amazon EMR Spark analytics operate under VeUP managed FinOps, while Amazon CloudFront handles global content delivery — the full media loop, from upload to safe playback, on AWS.

Production outcomes

KPI	Result
Safety gate quality	99.56% NSFW recall on the labeled benchmark (4 missed of 900 positives) — the moderation gate holding the platform’s “unsafe content never reaches users” editorial standard in production.
Media-pipeline economics	~57% lower moderation cost per 1,000 images versus the prior approach, with EKS Karpenter/Spot capacity and managed FinOps holding the wider compute estate’s cost curve.
Dual-path throughput	Synchronous image verdicts at interactive latency alongside an asynchronous video/bulk pipeline that scales per-frame fan-out on demand — one verdict contract across both paths and both engines.
Lessons & continuation	For UGC media platforms, trust and safety is a media-infrastructure problem: put a measured evaluation harness at the gate, keep an engine fallback behind a router, and run moderation, analytics, and delivery as one estate so quality, cost, and latency are managed together.

AWS services in production

Amazon Bedrock (Qwen 3 VL multimodal)Amazon Rekognition (fallback)AWS Step FunctionsAmazon EventBridge PipesAWS LambdaAmazon S3Amazon EKS (Karpenter/Spot)Amazon EMR (Spark)Amazon CloudFrontAmazon CloudWatch