Cluster Systems / ClusterPOS

Production OpenWebUI Retrieval-Augmented Generation chatbot on Amazon Bedrock for a European SaaS retail point-of-sale platform.

The customer

ClusterPOS is a European SaaS retail point-of-sale platform serving small-to-medium retail businesses. Their support workflow ran through a Zoho-hosted knowledge base — a static documentation tree with no conversational, grounded assistance — which meant linear support-headcount growth as the customer base expanded.

The challenge

ClusterPOS needed to deflect repeat support questions, accelerate time-to-resolution, and stay competitive against retail-POS rivals increasingly shipping AI-assisted in-product help. The team had no prior Generative AI capability in production and needed an AWS partner to take them from zero to a production chatbot within a sub-three-week build window.

The solution

VeUP designed and delivered a production OpenWebUI chatbot grounded on the customer's Zoho knowledge base, with Amazon Bedrock as the inference plane and Amazon Nova Pro as the production primary foundation model. The signed SOW preserved alternate-model-toggleable Bedrock integration so the customer could expand the model set post-launch without redeploying. Live POS data is reached through OpenWebUI tool/function-calling, letting end-users ask grounded questions over both static documentation and live operational data.

The production deployment runs in the customer's AWS Canada (Central) region (ca-central-1) across a Multi-AZ VPC, with Application Load Balancer + ACM TLS at ingress, ECS Fargate hosting the OpenWebUI and VeUP-forked Bedrock Access Gateway, and Amazon RDS for PostgreSQL with pgvector for application state plus the RAG index. Production cutover was confirmed 2025-11-25; a POS integration extension shipped to production 2026-05-13.

Architecture

AWS Well-Architected view of the ClusterPOS support agent — from the pre-engagement static-KB support workflow through to the production Multi-AZ VPC on Bedrock Nova Pro, with the RAG and live-POS tool surfaces, security, and observability planes.

Cluster Systems ClusterPOS — previous state. Static Zoho-hosted knowledge base with no conversational assistance, manual POS lookups against the live system, and no prior GenAI or agent workload (greenfield), with the self-host / data-residency constraint that ruled out third-party SaaS chatbot vendors. — Previous state — the pre-engagement support-operations baseline.

Cluster Systems ClusterPOS — target state on AWS. Multi-AZ VPC in ca-central-1 with public-subnet Application Load Balancer plus ACM TLS, private app-tier ECS Fargate task running OpenWebUI and a VeUP-forked Bedrock Access Gateway, private data-tier Amazon RDS PostgreSQL with pgvector plus Amazon EFS and AWS Backup, Amazon Bedrock (Nova Pro, Nova Lite, Claude Sonnet, Titan embeddings), AWS IAM, Secrets Manager, KMS, and CloudWatch, external Zoho KB and ClusterPOS POS API tool surfaces, and Google Workspace SSO. — Target state on AWS — production Multi-AZ architecture.

Cluster Systems ClusterPOS — layer-by-layer build-up of the target AWS architecture: network foundation, compute and application, data, security and observability, and the numbered data and network paths connecting them. — Layer-by-layer build-up — network foundation through data and network paths.

Cluster Systems ClusterPOS — AWS Well-Architected architecture diagram. Previous state: static Zoho-hosted knowledge base with no conversational assistance, manual POS lookups, and no prior GenAI or agent workload (greenfield). Target state: AWS Cloud in ca-central-1, Multi-AZ VPC with public-subnet Application Load Balancer plus ACM TLS, private app-tier ECS Fargate task running OpenWebUI and a VeUP-forked Bedrock Access Gateway, private data-tier Amazon RDS PostgreSQL with pgvector plus Amazon EFS, Amazon Bedrock (Nova Pro, Nova Lite, Claude Sonnet, Titan embeddings), AWS IAM, Secrets Manager, KMS, and Backup, external Zoho KB and ClusterPOS POS API tool surfaces, Google Workspace SSO, and Amazon CloudWatch observability. Includes a Well-Architected pillar-by-pillar alignment strip. — Full Well-Architected overview — VeUP-rendered from the VeUP engagement record and Agentic AI / Generative AI competency case-study evidence.

Production outcomes

KPI	Result
RAG retrieval latency	Sub-800 ms across a 600-document staging corpus; end-to-end response within 1–2 seconds — well inside the SOW SLA (p50 ≤ 5s, p95 ≤ 10s).
Production AWS run-rate	Amazon ECS Fargate ~$23/day after right-sizing (~$700/month, ~$8,400/year) sourced from AWS Cost Explorer / CUR — plus variable Amazon Bedrock token spend.
Customer AWS adoption signal	Customer-attested doubling of total AWS cloud bill post go-live (Fozail Ahmad, CTO, Final Handover 2025-11-25).
Continued engagement	Three downstream customer-initiated VeUP SA requests post-handover (RDS MySQL 8 / DDoS / WAF; IAM security review; NLB→ALB + WAF). Downstream Ignite managed-services attach Closed-Won 2025-12-09.

In the customer's words

“Based on everything we agreed to set out and achieve, I think we have done it. In fact, we have gone even above and beyond that with the semantic search layer which I'm really excited about.”
— Aries Youssefian, VeUP Solutions Architect, Final Handover summary 2025-11-25

“About 200 dollars a month right now... the 200 dollars is all the base infrastructure. And that gateway, this, that everything that's running.”
— Fozail Ahmad, CTO, Cluster Systems, Final Handover 2025-11-25 (customer-perceived base chatbot infra)

AWS services in production

Amazon Bedrock (Nova Pro primary, alternate-model-toggleable) · AWS Fargate (OpenWebUI + VeUP-forked Bedrock Access Gateway) · Amazon RDS for PostgreSQL with pgvector · Amazon EFS · Application Load Balancer · AWS Certificate Manager · AWS Secrets Manager · AWS IAM · AWS KMS · Amazon CloudWatch · AWS Cost Explorer. External integrations: Zoho KB ingestion · Google Workspace SSO (OIDC/SAML) · ClusterPOS POS API (tool/function-call).