This note confirms what the public Vinasoy materials support.
What is public
The source material supports these facts:
- the customer name: Vinasoy
- a hyper-personalized 1:1 customer-engagement program (Tet 2025) produced and delivered in production via Pixel ML's Agentic AI customer-engagement platform
- 300,000+ uniquely personalized customer interactions generated and delivered — every individual consumer received a one-to-one AI-generated experience
- users wrote a wish, chose a contractually licensed celebrity persona, and received a uniquely personalized AI video produced by an agentic workflow
- production deployment as a SaaS on AWS solution operated by Pixel ML using AgenticFlow on Amazon EKS, Amazon Bedrock, Amazon SageMaker, Amazon EC2 GPU, Amazon S3, and Amazon CloudFront
- the case maps to the Agentic AI Applications competency category — a multi-step orchestrated workflow performing per-customer planning, reasoning, tool use, and action execution on AWS, with Generative AI (TTS + lip-sync video generation) used as one tool inside the agentic chain
- production status (not pilot, not proof of concept) is confirmed by customer-approved campaign artifacts, AWS ACE-registered launched-stage opportunity records, and Pixel ML operational deployment records
What the campaign did
The campaign let users create a personalized Tet greeting for someone they cared about. Pixel ML built and operated the 1:1 AI video workflow behind that experience, with content-safety moderation on every input, contractually licensed celebrity personas, transparent AI-generated disclosure on every video, and end-to-end automation on AWS-managed services.
What stays private
These details are not on the public site and remain NDA-protected, available only to AWS for validation purposes:
- detailed cost-per-engagement and downstream-conversion metrics beyond the 300,000+ public output figure
- private campaign assets and internal asset library
- internal operating mechanics and scheduling logic
- persona-licensing economics
- proprietary fine-tuned model weights