AI with Guardrails

This boilerplate demonstrates how to run Qwen3-VL-8B locally using LM Studio with an OpenAI-compatible API, while implementing a powerful Guardrails system to protect your application against misuse, prompt injection, jailbreak attempts, and malicious actors.

It provides a complete, end-to-end example that combines local LLM development with a strong security layer — essential for any production-oriented AI workflow.

Overview

The example integrates Next.js App Router, shadcn/ui, and the Vercel AI SDK (AI Elements) to build a secure chat interface powered by a locally hosted model.

A well-structured guardrailSystemPrompt ensures the model follows strict behavioral boundaries, blocking attempts to bypass restrictions or extract sensitive information.

A companion file, PROMPT_TEST.md, is included to demonstrate real-world tests you can run against the guardrails — covering injection attempts, jailbreak scenarios, manipulation tactics, and malformed instructions.

This setup acts as a local AI security lab, allowing you to test and validate LLM guardrails before deploying them in production.

Features

LM Studio hosting Qwen3-VL-8B with OpenAI-compatible API
Guardrail system prompt designed to mitigate jailbreaks and malicious intent
Includes PROMPT_TEST.md with practical attack scenarios
Next.js App Router for modern routing
Vercel AI SDK (AI Elements) for a polished chat interface
shadcn/ui for consistent UI components
End-to-end chat with enforced safety boundaries
Fully local, cost-free LLM experimentation
Ideal for testing safe AI implementation before deploying to cloud models

Use Case

Perfect for developers who want to:

Experiment with LLM security in a controlled, local environment
Validate guardrails against real attack patterns
Build AI applications where safety and reliability are critical
Test local LLMs before moving to hosted providers
Run a fully offline development workflow without token cost

With LM Studio, Next.js, AI Elements, and a robust guardrail system, this boilerplate offers a practical, security-focused foundation for building AI applications that remain safe — even under adversarial pressure. 🚀

Boilerplate