How LocaleFrame works.

Five layers, each run on a provider chosen for that job, behind one control plane on AWS. Each layer scales on its own and can be replaced without touching the others.

Select a layer to see what it does.

Orchestration & Async Queue

AWS ECS · SQS · RDS · S3

Control plane. Owns multi-tenant campaign state and drives every job through the pipeline.

API gateway, auth and per-workspace rate limiting on ECS Fargate
Campaign / variant state machine in RDS PostgreSQL
SQS FIFO queues per stage with dead-letter redrive
Enterprise asset vault in S3 with per-tenant KMS keys
Signed webhook dispatch with exponential back-off

Interfaces

REST /v1/*
SQS: lf-script, lf-render, lf-score, lf-mux
S3 presigned URLs

The layers

Layer	Runs on	Job
Orchestration & Async QueueAWS ECS · SQS · RDS · S3	AWS ECS · SQS · RDS · S3	Control plane. Owns multi-tenant campaign state and drives every job through the pipeline.
Cultural Intelligence & ScriptingAnthropic Claude API	Anthropic Claude API	Turns a product page into culturally native creative for 35+ locales.
Generative Motion & B-RollRunway Gen-4.5 · Act-One	Runway Gen-4.5 · Act-One	Produces cinematic motion from static SKU photography.
Retention Scoring & TelemetryGoogle Cloud Vertex AI · BigQuery	Google Cloud Vertex AI · BigQuery	Quality gate before media spend.
FFmpeg Muxing & Global DeliveryAkamai Connected Cloud (Linode) GPU · Edge CDN	Akamai Connected Cloud (Linode) GPU · Edge CDN	Final assembly and worldwide streaming.

Building on it? The API reference covers the render endpoint and lifecycle webhooks.

One render, start to finish.

Target timings for one product rendered for three markets.

T+0.0s
AWS. POST /v1/campaigns/render accepted, idempotency key stored in RDS, job fanned out per locale to SQS.
T+1.2s
Claude. Product page scraped; brand voice and value props extracted; hooks, VO and subtitles transcreated per locale.
T+9s
Runway. Keyframes + camera prompts submitted; 4 shots synthesized per variant with regional lighting.
T+62s
Vertex AI. Saliency and 3s retention scored; variants below threshold are re-scripted and re-rendered.
T+71s
Akamai. FFmpeg NVENC mux, subtitle burn-in, HLS ladder packaging, edge publish and cache warm.
T+84s
AWS. delivery.live webhook signed and dispatched; campaign state finalized.

Variant states

variant.status

queued ─► scripting ─► rendering ─► scoring ──┬─► muxing ─► live
                                              │
                         (score < threshold)  └─► scripting (max 2 loops)
any stage ─► failed (after 3 retries, quota refunded)

Queue-isolated stages

Each engine consumes from its own SQS queue with independent concurrency and dead-letter redrive, so a slow model provider never blocks another stage.

Idempotent, resumable jobs

Every stage writes content-addressed artifacts to S3 before acknowledging. Retries resume from the last completed artifact instead of re-running paid inference.

Tenant isolation

Per-workspace KMS keys, row-level security in RDS and scoped presigned URLs. Brand glossaries and voice references never leave the tenant boundary.

Egress-aware delivery

Final renders are written to Linode Object Storage and served from Akamai's edge, keeping high-volume video traffic off hyperscaler egress.