Industry Case Study

Secure, low-cost LLM support chatbots for small business deployments

Explore the Architecture, Platform design, Delivery Pipeline, and Findings from a k3s-based edge AI private cloud with DGX unified memory and multi-tenant isolation.

Industry Case Focus

Chatbot workflow showing retrieval, guardrails, and response assembly for industry support teams.

Chatbot workflow diagram

Architecture

Edge AI private cloud.

Edge AI computing cluster architecture diagram

Edge AI computing cluster built on k3s with unified DGX nodes.

Platform

Architecture highlights

Distributed k3s clusters

Lightweight Kubernetes nodes span heterogeneous, low-cost machines while maintaining a consistent control plane.

Encrypted overlay network

Overlay routing connects nodes securely, enabling pooled compute without centralizing all infrastructure.

Multi-tenant isolation

Container-based tenancy boundaries and per-tenant data access controls keep client data separated and auditable.

No-code workflow

Small business operators can deploy and configure chatbots without dedicated ML engineering teams.

Delivery Pipeline

From code submission to live deployment

💻
Code submit

Feature branches land in the mainline.

Quality checks

Linting, tests, and security scans.

🏷️
Release version

Semantic tagging for traceability.

📦
Build image

Container build and registry publish.

🚀
Deploy to cluster

Automated rollout to edge AI nodes.

Findings

Real-world deployment findings

E-commerce validation

The platform is tested in a live customer support setting, demonstrating stable performance under realistic traffic.

Operational constraints

Resource pooling and lightweight clusters deliver lower total cost while preserving reliability.

Security impact

Practical defense layers reduce prompt injection exposure without retraining or specialized hardware.

Authors

Jiazhu Xie [1]
S4076491@student.rmit.edu.au
RMIT University, Melbourne, VIC, Australia
Bowen Li [1]
S3890442@student.rmit.edu.au
RMIT University, Melbourne, VIC, Australia
Heyu Fu
S4153648@student.rmit.edu.au
RMIT University, Melbourne, VIC, Australia
Chong Gao
cyrus.gao@rmit.edu.au
RMIT University, Melbourne, VIC, Australia
Ziqi Xu
ziqi.xu@rmit.edu.au
RMIT University, Melbourne, VIC, Australia
Fengling Han
fengling.han@rmit.edu.au
RMIT University, Melbourne, VIC, Australia

[1] These authors contributed equally to this work.