ClusterPOS empowers its support team with an AI agent on Amazon Bedrock

For an ISV, adding AI to a live multi-tenant product is a different problem from standing up a chatbot. The assistant had to run inside ClusterPOS’s SaaS operating model: predictable unit economics on shared infrastructure, tenant data kept clean, credentials and integrations (the Zoho knowledge base, and later live POS data) managed safely, and the whole feature held to the same reliability bar as checkout — because a POS vendor’s customers do not distinguish “the AI is down” from “the product is down.” ClusterPOS needed the feature shipped to production with acceptance criteria it could hold its own platform to.

VeUP designed and delivered the assistant as SaaS infrastructure. The serving layer runs on Amazon ECS Fargate behind an Application Load Balancer with AWS Certificate Manager TLS — no instances to manage, capacity scaled to demand, CPU reservations measured and right-sized after launch. Amazon Bedrock (Amazon Nova Pro) provides the model layer with no keys or GPU fleet to operate; Amazon RDS for PostgreSQL and Amazon EFS carry application state; AWS Secrets Manager isolates integration credentials; AWS KMS encrypts at rest. The build was held to explicit acceptance criteria and proven at a formal gate: latency percentile targets under nominal load, zero critical CloudWatch alarms, grounded answers citing their sources from the seeded knowledge base, and a backup-restore test into staging passed via AWS Backup. The engagement continued past launch into solution-architecture support as the assistant’s scope grew.