Why we constructed the Safe AI Manufacturing facility
One 12 months in the past at NVIDIA GTC, Cisco and NVIDIA launched the Cisco Safe AI Manufacturing facility with NVIDIA. On the time, the problem for enterprises was clear: AI was transferring from science venture to strategic precedence, however the infrastructure and software program have been fragmented. Clients have been battling the complexity of standing up huge compute clusters for coaching, optimization, and inferencing, whereas guaranteeing their knowledge remained safe and personal.
To unravel this, the Safe AI Manufacturing facility was constructed upon the inspiration of Cisco AI PODsour modular reference structure based mostly on Cisco Validated Designs and NVIDIA Enterprise Reference Structure–compliant deployments. By integrating Cisco networking, compute, and accomplice storage with NVIDIA’s accelerated infrastructure, the full-stack AI POD permits enterprises to seamlessly deploy AI purposes with enterprise-grade safety fused into the material. Embedded into every layer of the reference structure and wrapped round it’s the Cisco portfolio of safety and observability capabilities options. It offers companies with a dependable, safe, hardened path to coaching, optimization, and inferencing throughout the core knowledge heart.
The rise of multi-agents all over the place
Within the final twelve months, the panorama of enterprise AI has shifted. We’ve got moved past easy generative AI chatbots and have now transitioned into the period of autonomous brokers that drive business-critical features. The main driver right now is the necessity to run these multi-agent techniques all over the place: within the core, within the cloud, and more and more, on the edge.
Nonetheless, this shift has revealed a large hurdle: Safety is at present the one greatest barrier to the enterprise adoption of AI brokers.
In contrast to a regular chatbot, an agent is autonomous; it has the ability to invoke APIs, entry knowledge lakes, and make selections, thus enabling groups to lower time to worth. These brokers depend on massive language fashions (LLMs) and small language fashions (SLMs) to offer reasoning context. If the mannequin offering that context is attacked or goes down, the affect is catastrophic:
- Compromised operational resolution making: If an agent managing warehouse logistics is hijacked via immediate injection, it might be manipulated to authorize incorrect delivery priorities or “ghost” stock, resulting in vital margin erosion and unfulfilled buyer orders.
- Crucial workflow stagnation: In a high-velocity surroundings, if an edge mannequin goes down or its reasoning is corrupted, the automated course of stops. For a warehouse, this implies vans aren’t loaded and the availability chain grinds to a halt, straight impacting the quarterly backside line.
- Regulatory and compliance publicity: Brokers working on the edge usually deal with delicate knowledge. A safety breach that results in the exfiltration of personally identifiable data (PII) or proprietary predictive fashions can lead to huge regulatory fines and a everlasting lack of buyer belief.
Extending Safe AI Manufacturing facility throughout the core to the sting
To satisfy these challenges, Cisco is saying that it’s increasing Cisco Safe AI Manufacturing facility from the info heart to the sting, guaranteeing safety follows the agent throughout your complete multi-agent spectrum.
With Cisco AI Protection, we offer a constant safety posture no matter location. Whether or not an agent is working a trillion-parameter LLM within the knowledge heart or a specialised SLM in a distant warehouse, Cisco ensures the mannequin is validated and the prompts are sanitized in actual time.
A brand new reference structure: Predicting stockouts for retail or manufacturing warehouses with Cisco Safe AI Manufacturing facility with NVIDIA and its ISV ecosystem
Whereas the Cisco AI POD offers a high-performance engine within the knowledge heart, a brand new problem for some organizations lies on the warehouse flooring. At GTC, we’re showcasing how a warehousing workforce can leverage Cisco and accomplice structure to search out worth via visibility: understanding the precise second a stockout happens or a security hazard arises earlier than it impacts the morning shift. For IT, the problem is bridging the IT/OT hole by deploying and securing AI throughout lots of of distant websites with out making a administration nightmare. Extending visibility and operations from the core knowledge heart to the sting is about extra than simply {hardware}; it’s about getting the fitting intelligence to the fitting individual, securely and at scale.
At GTC, we’re demonstrating this growth via a real-world resolution extending the NVIDIA Multi-Agent Clever Warehouse blueprint. This demo showcases how specialised brokers can bridge the hole between IT and OT layers.
How the answer elements coordinate
The workflow begins on the edge, the place Vaidio acts because the “eyes” of the warehouse, monitoring video feeds for occasions like pallet stockouts. When a stockout or low inventory occasion is detected, Vaidio triggers a REST API name to an Aible agent working domestically on Cisco Unified Edge. This agent makes use of an SLM to offer quick reasoning context—deciding if a stockout is crucial sufficient to disrupt a pending cargo. If motion is required, the sting agent pings a core agent within the Cisco AI POD hosted within the knowledge heart, which queries the enterprise knowledge lake and LLM to calculate the income affect.
For the warehouse supervisor, this closes the IT/OT hole totally. As an alternative of manually cross-referencing spreadsheets or discovering empty cabinets too late, the supervisor has a 24/7 digital assistant that identifies issues, calculates the enterprise price, and triggers an expedited order routinely, eliminating “ghost stock” and preserving the availability chain transferring.
The answer structure
- Vaidio (the eyes): Operating on Cisco Unified Edge nodes (3-4 nodes per web site), Vaidio pc imaginative and prescient containers make the most of NVIDIA L4 GPUs to observe the ground for stockouts or security hazards.
- Aible (the mind): An agentic platform orchestrating the workflow throughout the sting and core. When Vaidio detects a problem via a RESTful API, an Aible agent on the edge makes use of a neighborhood SLM (equivalent to Nemotron-3 Nano or Llama 3.3) to offer quick reasoning context.
- Cisco AI POD (the core): For heavy-duty predictive modeling, edge brokers talk again to the core AI POD (powered by NVIDIA GPUs). This core long-running agent powered by Aible understands find out how to stability the affect of stockouts in opposition to the price of expedited delivery, queries the warehouse knowledge lake to take the fitting selections based mostly on income affect and invoke delivery APIs.
- The deployment bundle: This resolution runs Kubernetes 1.34 on Ubuntu 24.04 LTS with NVIDIA Driver 580.95.05 throughout each edge nodes and core AI POD infrastructure.
Cisco Intersight offers centralized AI server fleet administration throughout core and edge, delivering infrastructure lifecycle management, coverage consistency, configuration governance, and operational visibility throughout distributed AI environments.
This ensures the underlying compute platform stays constant, safe, and scalable, whereas utility and container orchestration function inside a validated, enterprise-ready infrastructure basis.
Cisco AI Protection and complete safety for AI brokers and purposes
There’s a brand new actuality within the enterprise: multi-agent techniques. And Cisco is the one vendor that embeds safety into the material of the AI Manufacturing facility, defending every part from the infrastructure to the brokers that stay on it.
Cisco AI Protection offers a dual-layer defend from the info heart to the sting:
- Mannequin integrity and scanning: AI Protection scans the LLMs and SLMs via their uncovered APIs/endpoints. This identifies potential vulnerabilities in mannequin recordsdata and identifies the “AI Invoice of Supplies” (AI-BOM) to make sure provide chain integrity.
- Runtime guardrails: As brokers talk, AI Protection applies real-time insurance policies to sanitize prompts and responses. It detects immediate injections, prevents the technology of poisonous content material, and ensures that delicate knowledge—together with PII, PHI, and PCI—by no means leaves the safe surroundings.
- Platform safety: Cisco Unified Edge provides a bodily and firmware layer of safety, together with firmware roots of belief, locking bezels with intrusion detection, and Intel TDX/SGX for confidential computing.
The Aible distributed agent resolution highlighted above advantages from all three layers of safety. It runs securely on Cisco Unified Edge and Cisco AI POD servers. The fashions utilized by it on the core and the sting are scanned by AI Protection. All language mannequin interactions of the brokers on the core and edge use the runtime guardrails by default.
The way forward for multi-agents all over the place
The transfer to multi-agent AI represents the following industrial revolution of intelligence. By fixing the challenges of edge deployment and mannequin safety, Cisco helps clients transfer previous the experimental section and into real-world manufacturing.
By bringing the NVIDIA Multi-Agent Clever Warehouse blueprint to life on Cisco structure, we’re lastly delivering the “chocolate” to either side of the enterprise. The Warehouse Supervisor positive factors a 24/7 digital coordinator via Vaidio and Aible that automates crucial workflows and eliminates “ghost stock.” Concurrently, the IT worker positive factors a safe, “drift-free” platform managed with Cisco Intersight and guarded by AI Protection. This isn’t only a technical achievement; it’s a production-ready resolution that turns AI potential right into a aggressive benefit for your complete enterprise.
Keep tuned for our upcoming technical white paper, which is able to present a deep dive into the structure of the Safe Multi-Agent AI Manufacturing facility.
Able to speed up your time to worth? Contact your Cisco account consultant right now to learn to lengthen your safe AI technique from the core to the sting or click on right here to study extra in regards to the Cisco Safe AI Manufacturing facility with NVIDIA.
