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 BedrockAWS Step FunctionsAWS LambdaAmazon S3

For a global consumer social platform, VeUP built content moderation as an orchestrated multi-model system rather than a single model call: a Qwen 3 VL classifier-with-policy-prompt on Amazon Bedrock does the reasoning, an engine router abstracts it from an Amazon Rekognition fallback, a K-of-N aggregator turns per-frame judgments into a single video verdict, and AWS Step Functions orchestrates the whole flow with a Map state throttled at MaxConcurrency=5.

The challenge

The platform’s safety policy is asymmetric: unsafe content must never reach users, while over-blocking erodes the product. Meeting that bar with a single moderation model is brittle — models disagree frame to frame on video, no one engine is best forever, and a hard dependency on any single engine leaves no safe path to swap or fall back. The engineering problem was agentic in nature: how to compose a policy-reasoning model, an alternate engine, and a deterministic decision rule into one system that renders exactly one defensible verdict per piece of content, at platform scale, without melting the inference tier.

The solution

VeUP structured the pipeline as a chain of specialized roles. The classifier is Qwen 3 VL on Amazon Bedrock, prompted with the platform’s written moderation policy — the model reasons against the actual policy text rather than a fixed label taxonomy. It was selected for production after a measured Phase 1 head-to-head in which Llama 4 Maverick, Llama Guard 4, Amazon Nova Premier, and Claude Opus were evaluated as candidates on a 1,501-image labeled benchmark. An engine-router abstraction presents one verdict contract over the Bedrock engine and the Amazon Rekognition fallback, so engines can be swapped by configuration, not re-architecture. For video, frames are extracted and judged independently, then a K-of-N aggregation rule composes the per-frame outputs into a single verdict — deterministic logic bounding probabilistic components. AWS Step Functions runs the per-frame fan-out as a Map state with MaxConcurrency=5, deliberately throttling Bedrock invocations so throughput scales without tripping service limits, with AWS Lambda doing the work at each step and Amazon S3 carrying media through the flow.

Production outcomes

KPI	Result
Verdict quality	99.56% NSFW recall on the 1,501-image labeled benchmark — the decisive metric under the platform’s “false negatives never” policy — carried into Phase 2 production.
Orchestration economics	~57% lower inference cost per 1,000 images versus the prior engine, with the Step Functions Map throttle keeping Bedrock invocation rates inside service limits at platform scale.
Composability	One verdict contract across two engines: the router lets the platform change its primary model or fail over to Rekognition without touching the pipeline — the head-to-head evaluation harness remains standing to re-run as new models appear.
Lessons & continuation	Agentic systems earn trust through deterministic composition: policy-prompted reasoning where judgment is needed, K-of-N rules where consistency is needed, and explicit concurrency control where scale is needed. The router-plus-harness pattern turns model churn from a risk into an option.

AWS services in production

Amazon Bedrock (Qwen 3 VL)Amazon Rekognition (fallback engine)AWS Step Functions (Map, MaxConcurrency=5)AWS LambdaAmazon S3Amazon CloudWatchAWS IAM