AI
The Problem: Too Many Experts, Too Much Complexity
Deploying a 3GPP network today requires five to ten separate specialists: one for the core, one for the RAN, one for Kubernetes, one for AI, and so on. Specifications are dense and scattered. Infrastructure built on ONAP is expensive. Vendors push proprietary lock-in. Ariza’s diagnosis: the industry builds cathedrals when it needs clinics. The patient is overweight, over-engineered, and over-budget.
One CRD, One YAML File, Any Cellular Network
Ariza built a Kubernetes CRD he calls a universal 3GPP cell network blueprint. A single YAML file specifies the generation (4G or 5G), the core software, and the RAN software. One Helm chart deploys everything. The CRD maps to CNCF tooling at each lifecycle stage: Kind and Helm for day-zero provisioning, Flux for day-one GitOps reconciliation, Grafana and Prometheus for day-two observability, and Keptn for self-healing. Today it covers 5G standalone and LTE. NB-IoT, 5G broadcast, and LTM are planned.
Kim: A 3GPP AI Expert Grounded in Specs
The AI assistant is named Kim. It connects through an MCP server to the CLI, so it can only call commands Ariza explicitly exposes. Kim answers questions about 3GPP by querying a RAG index built from the testspec-llm repository, loaded into an Obsidian vault. Supported models include Mistral, Qwen, Llama, and NVIDIA’s Neotron, which launched at Mobile World Congress targeting telco use cases. Deutsche Telekom and Google announced a strategic partnership to build similar network AI. Ariza’s constraint is firm: the CLI orchestrates everything, Kim only assists.
Four Live Demos: From Slicing to Voice Deployment
Demo one deployed a full 5G private network end-to-end from Kind cluster creation through Grafana, Prometheus, and Keptn installation to a working smartphone connection. Demo two showed Kim catching a deliberate error: Ariza asked about “4G NSA,” which does not exist, to verify Kim flags bad input. Demo three used Open Air Interface to run network slicing, configured by answering CLI prompts and visualized in real time as distinct slices in Grafana. Demo four had a staff member with no Kubernetes background speak a voice command into Claude, which routed through MCP to the CLI and deployed a 5G network named “Annab breath” on band 78.
Hardware Floor and Deployment Paths
The Docker Compose path runs on 8 GB of RAM with an 8-core 2.6 GHz processor. Adding Flux, Grafana, Prometheus, and Keptn requires 16 GB of RAM at the same core count. Running a local LLM adds a GPU requirement. Ariza’s demo machine carries a GPU with 16 GB of VRAM to handle the billions of parameters needed for the local chatbot. Remote API models sidestep that requirement for teams without GPU hardware.
Q&A
What hardware is needed to run the demos? Docker Compose 5G needs 8 GB RAM and 8 cores at 2.6 GHz; the full Kubernetes stack with observability needs 16 GB RAM; local LLM inference needs a GPU with at least 16 GB VRAM. ▶ 27:58
Can the PDN/PDP context be preserved across redeployments so the phone does not need to reconnect? Ariza said the demo used the laptop’s wired or wireless connection as the uplink, not a full GTPU tunnel to the phone, so the question did not apply to his current setup. ▶ 30:29
Notable Quotes
the industry built cathedrals when it needs clinics. Luis Ariza · ▶ 04:19
the CLI orchestrate everything but the AI only assist and never takes the wheel. Luis Ariza · ▶ 11:30
So the guy doesn’t need a PhD to execute a private cellular network. It’s pretty easy. Luis Ariza · ▶ 23:34
from a five year old to 120 years old who can deploy a 5G private cellular network your own cellular network is one conversation away. Luis Ariza · ▶ 26:11
Key Takeaways
- One YAML CRD replaces five-to-ten specialist roles for 3GPP network deployment.
- Kim’s AI answers are grounded in a RAG index of the full 3GPP specification set.
- The Docker Compose path needs only 8 GB RAM and 8 cores to run 5G standalone.
About the Speaker(s)
Luis Ariza holds a PhD in telecommunications and focuses on Open RAN, cloud-native 5G, NTN, IoT, and private networks. He is a Kubestronaut, organizer of the Colombia OpenInfra User Group and KCD Colombia 2025, and founder of Beanters. He has tracked Kubernetes-in-telco since 2019, when he built his first 5G private network on Kubernetes with professor Raymond.