DEMO

Real-Time AI Guardrails

Detect unsafe inputs, hallucinated answers and low-quality responses as they happen, and block them before they reach your customers, with configurable guardrails.

Building Blocks used:

ai.govern

Skills used

Download skills for IBM Bob →

real-time-guardrails

Products Used

1 Review use case

As AI agents become part of everyday business operations, one question becomes critical: how do you make sure they are safe, reliable and delivering the right results? A user might share sensitive information or try to bypass security controls, and the AI might produce an answer that sounds convincing but is not true. Real-Time AI Guardrails, powered by IBM watsonx.governance, detects these risks as they happen and enforces your policies before they affect your business. This demo shows how, in a web application you can run, tune and rerun in seconds.

Different risks need different protection, so the demo draws on a catalog of 18 guardrail metrics in three families. Content safety metrics detect personal information exposure, harmful content, jailbreak attempts and bias. RAG metrics check whether a generated answer is relevant and grounded in the retrieved documents, so unsupported claims are stopped before they reach a customer. Response quality metrics use an LLM as a judge to assess completeness, conciseness and helpfulness, and a custom Action Oriented Validator shows how to encode a rule of your own, such as requiring every troubleshooting answer to include next steps. Each metric has its own configurable threshold, so a customer service agent that must collect personal data can be tuned differently from one that must never see it.

Four scenarios walk through the metric categories one at a time: an input that combines personal data with a jailbreak attempt, an answer that claims facts its sources never mention, a long response that leaves out essential steps, and an explanation that never tells the customer what to do next. In each case you see which guardrails fired, the scores behind the decision, and the request or response being blocked. Because the guardrails are built with the watsonx.governance SDK, the same calls run inside your own agents, applications and workflows, with configurable policies and audit logging.

Start Here

Watch this short overview

Preview the application

Explore key screens from the demo

2 Try the interactive application

Experience the interactive demo and explore its features. A demo script is provided to help when presenting the application to partners.

Try the demo

Have questions, want a walkthrough or found a bug?

Contact the demo owner for assistance.

Submit a request →

3 Explore how it was built

The architecture has three layers. A web application on IBM Cloud Code Engine is the orchestration layer: it takes the user's selected guardrails and thresholds, sends one batched evaluation request, and turns the returned scores into a Pass or Block decision per metric. IBM watsonx.governance is the evaluation engine, where Granite Guardian models score the content safety and grounding metrics. IBM watsonx.ai supplies the LLM-as-judge for the quality metrics, calling Llama 3.3 70B. A custom metric built with the watsonx.governance SDK shows how organizations can extend the catalogue with their own rules.

Architecture

4 Download and modify with IBM Bob

Download the source code and specifications for this application to set up and run it locally. Once downloaded, open the code in IBM Bob to modify and adapt it to your own use case.

Source Code

Real-Time AI Guardrails — Seller Demo Script

Before you start: Open the demo URL. The left column (Select Guardrails) lists metrics in three groups, each with a checkbox and a Threshold field. Above the query are four scenario buttons. The query, context and response fields are read-only. Scenario buttons fill the text — you tick the metrics yourself.

1. Set the scene

Key message: A user may share sensitive data or try to bypass controls, and the model may answer convincingly but wrongly. Real-Time AI Guardrails, powered by watsonx.governance, catches this as it happens.

As AI agents join everyday operations, the question is how to keep them safe, reliable and correct.

2. Tour the catalog

Key message: A single safety check is not enough — each risk gets its own metric, and each metric has its own threshold.

Point at the three groups on the left: Content Safety, RAG, and Response Quality — 18 metrics in all.

3. Scenario 1 — Unsafe input

Key message: The request is blocked before it reaches the model. A customer service agent that needs to collect personal data can tune the PII threshold differently.

Click Content Safety.
Tick PII Detection, Harm Detection, and Jailbreak Detection.
Click Run Guardrails.
Show the High Risk tile and the flagged rows in the table.

4. Tune a threshold

Key message: For different use cases, you will need different thresholds.

Change a Threshold field — e.g. raise PII Detection from 0.65 to 0.95.
Click Run Guardrails again and show the result flip.

5. Scenario 2 — Unsupported answer (hallucination)

Key message: The answer claims facts the reference documents never mention. This stops a hallucination before it reaches a customer.

Click RAG: Unsupported Answer.
Keep Jailbreak Detection ticked; add Answer Relevance, Context Relevance, and Faithfulness.
The context and generated response fields open.
Click Run Guardrails: Faithfulness blocks while the two relevance metrics pass.

6. Scenario 3 — Safe but incomplete

Key message: These use an LLM as a judge — you enforce the quality of the experience, not only safety and accuracy.

Click Response Quality: Long & Incomplete.
Add Answer Completeness and Conciseness.
Click Run Guardrails: both block.

7. Scenario 4 — A rule of your own

Key message: You can encode your own requirements using the same SDK — no custom infrastructure required.

Click Response Quality: No Next Steps.
Keep Helpfulness and add Action Oriented Validator.
Click Run Guardrails: the custom validator blocks.
Explain: the policy requires actionable next steps in every troubleshooting answer, and this validator was written with the same watsonx.governance SDK.

8. Close

Key message: Built-in and custom guardrails, configurable thresholds, policies enforced in real time. The same SDK calls run inside your own agents, applications and workflows, with audit logging.

Mention the source code is in the repository for reference.