## **Demonstrating Secure Confidential AI for HR Agentic Experiences**

In today’s enterprise landscape, using AI to automate complex workflows involving sensitive data requires more than just performance — it demands verifiable trust, security, and policy enforcement. In a recent demo, Jamie Aliperti, Director of Solutions at OPAQUE, showcased how our Confidential AI platform enables a secure, end-to-end agentic RAG pipeline, using an internal HR application as the use case.

Confidential AI Solutions for Secure Data Management 🔒 - YouTube

[Confidential AI Solutions for Secure Data Management 🔒](https://www.youtube.com/watch?v=I01IWA8wam8) Opaque Systems, Inc.384 subscribers

### **A Secure Agentic RAG Workflow, Built for Confidential Data**

The demo walked through a complete RAG (Retrieval-Augmented Generation) pipeline designed to answer employee HR questions across a large organization. The key technical building blocks include:

- **End-to-End Encrypted Data Sources**: HR documents and internal data remain encrypted throughout their lifecycle. Only the data owner retains the decryption key, and the data is never exposed — not even during model inference.
- **Hosted Models and Agents**: The pipeline includes a hosted [**IBM Granite**](https://www.ibm.com/granite?utm_content=SRCWW&p1=Search&p4=43700081742394771&p5=e&p9=58700008798092302&&msclkid=97aa2b15d9bf18001c8d439788b78bb9&gclid=97aa2b15d9bf18001c8d439788b78bb9&gclsrc=3p.ds&gad_source=7) model and multiple specialized agents orchestrated in a secure chat workspace. The agents manage task routing and contextual enrichment.
- **NeMo Guardrails for Policy Enforcement**: Integrated [**NVIDIA NeMo Guardrails**](https://developer.nvidia.com/nemo-guardrails?sortBy=developer_learning_library%2Fsort%2Ffeatured_in.nemo_guardrails%3Adesc%2Ctitle%3Aasc&hitsPerPage=12) enforce predefined rules, ensuring that only permitted questions are processed — protecting against prompt injection, out-of-scope queries, or leakage of sensitive responses.
- **Attested Preflight Verification**: Before any interaction begins, OPAQUE performs a trust attestation — a preflight check that verifies each asset (data source, model, agent) is authorized and will interact in only permissible ways.
- **Confidential Compute Runtime**: All processing happens inside a trusted execution environment using hardware-based confidential computing, ensuring runtime data protection.

### **Compliance Built-In**

After each run, the system produces a **tamper-proof audit log** cryptographically signed by the underlying CPU or GPU. These logs can be exported and verified using third-party tools — supporting both internal governance and external compliance requirements.

A visual **trust screen** makes the entire flow transparent: from attestation to audit, organizations can see which assets were used, how data moved, and validate that every action followed policy.

### **Built to Power Sensitive Workflows at Scale**

While the demo focused on an HR Q&A scenario, the architecture is adaptable to any sensitive workflow — whether in finance, healthcare, legal, or enterprise operations. With support for tools like [**LandGraph**](https://www.langchain.com/langgraph) for enriched data mapping and semantics, the platform supports complex enterprise-grade deployments.

OPAQUE’s Confidential AI platform enables teams to securely orchestrate GenAI pipelines with hosted LLMs, policy-bound agents, and protected data — all while maintaining visibility and control.

**This is what it means to build with AI you can prove.**