EdgeForge AI: From Prototype to Product — The Story of How We're Redefining Cloud Provisioning

EdgeForge AI: From Prototype to Product — The Story of How We're Redefining Cloud Provisioning

Every product is born from a real problem. EdgeForge AI is no exception.

The problem we wanted to solve

Anyone who works with open-source cloud infrastructure knows it well: provisioning is one of the most repetitive, time-consuming, and human-error-prone tasks in the entire infrastructure lifecycle.

Need a new VM with specific software? Open the terminal, write the Ansible playbook, configure the variables, test, fix errors, redeploy. For a simple environment: hours. For a complex environment with multiple VMs, dedicated networking, and interdependent software: days.

All this with the constant risk of configuration errors appearing in production at the worst possible time.

At Epic Edge, we've managed and continue to manage OpenStack infrastructure for some of Europe's largest organizations. We've faced this problem every day, for years. At some point, we decided to stop suffering and build a solution.

The idea: what if it were enough to describe it?

The starting point was as simple as it was ambitious: what if an engineer could describe the environment he wanted in natural language, and have it ready without writing a single line of code?

Not a wizard with static forms. Not a rigid template to fill out. A system that understands the intent, designs the architecture, generates the code, and executes it—adapting to the specific context of each infrastructure.

This is the idea that gave birth to EdgeForge AI.

Development: Months of work to make AI reliable in production

Building an AI system that works in a test environment is relatively simple. Building one that runs in production, on real enterprise customer infrastructures, with predictable and reliable results, is another matter entirely.

The main challenge wasn't technical in the traditional sense. Existing AI models are capable. The challenge was to make them reason deterministically—that is, to ensure that, when faced with the same request, the system always produced consistent and correct architectural choices, not variable outputs dependent on the model's "mood."

We built a pipeline with two AI models working in parallel—one specialized in natural language interpretation, one in code generation—orchestrated by precise instructions that define how the two models should collaborate and what constraints they should respect.

The result is a system that reasons, not one that guesses.

We've added a memory mechanism: successfully generated Ansible playbooks are archived. The next time the same software is requested, the system doesn't regenerate the playbook from scratch but uses the previously validated one—reducing time and increasing reliability. If the request is new and the system doesn't have enough information, it searches online, analyzes the documentation, and generates the correct playbook.

All of this runs completely on-premise. The AI models are installed on the customer's infrastructure, requests never leave the company perimeter, and no data is sent to external cloud services.

We Make Future 2026: Our First Public Stage

In June 2026, we were at We Make Future in Bologna—one of the most important tech fairs in Italy—with the lite version of EdgeForge AI.

Lite doesn't mean incomplete. It means we deliberately chose to present the core component of the product: text chat. The user writes what they want in natural language, the system interprets the request, presents a detailed deployment plan with the number of operations, the resources that will be created, and the estimated time, asks for confirmation, and then executes everything automatically.

A real-world example we showed at WMF: a full-featured monitoring environment with Grafana, Prometheus, Loki, and node_exporter on 5 VMs — 11 operations executed in under 25 minutes, with the environment fully functional, SSL configured, and login credentials ready at the end.

The audience's response confirmed what we had intuited: the problem EdgeForge AI solves is real and deeply felt. The engineers who stopped us at the booth weren't asking if it was useful—they were asking when they could have it.

What's coming: the interactive section

While we were preparing the lite version for WMF, in parallel — and for months — we were developing the next part of EdgeForge AI: the interactive section.

This is a graphical interface integrated directly into OpenStack Skyline, the OpenStack management dashboard. Alongside the text chat, the user finds an area organized into macro-categories that allows them to perform the most common operations without having to type anything.

There are five macro-categories:

🖥️ Compute — VM creation and management, SSH keypair

☸️ Kubernetes — K3s clusters, highly available K8s, custom configurations

💾 Storage — volumes, snapshots, backups

🌐 Networking — networks, subnets, security groups, floating IP, ports

🔐 VPN & Security — Site-to-site VPN

The choice to organize the features into macro-categories instead of displaying all the buttons at once is no coincidence. With over twenty operations available, a flat catalog would be confusing. Five clear categories make the interface accessible even to those who aren't OpenStack experts.

The underlying logic remains unchanged from chat: when the user selects an action and fills in the parameters, the system runs the same AI pipelines, generates the same dynamic Ansible playbooks, and reports the results back to chat when all jobs are complete.

The advantage of the interactive section is accessibility: you don't need to know how to describe the infrastructure in natural language to use it. A junior operator can create a VM, configure a security group, or deploy a Kubernetes cluster in just a few clicks, with the same reliability as a chat deployment by a senior engineer.

Why we chose to present the dispute to the WMF and not the full version

The answer is simple: we wanted real feedback before releasing the full version.

Text chat is the heart of the product—it's the part that demonstrates the core value of EdgeForge AI: AI that understands what you want and builds it. The interactive section is an extension that enhances accessibility but doesn't change the value proposition.

Presenting the solution allowed us to validate the most important thing—that the problem is real and that our solution actually solves it—before investing the final weeks of development to bring the interactive section to the level of quality we believe is necessary for a production release.

The feedback gathered at the WMF has been incorporated. The full release is imminent.

What this means for open-source cloud infrastructure

EdgeForge AI isn't a tool to simplify the cloud for those unfamiliar with it. It's a tool to free engineers from repetitive work and allow them to focus on what creates real value: architecture, optimization, and innovation.

Manual provisioning of cloud infrastructure will always be necessary for complex and non-standard scenarios. But the vast majority of day-to-day operations—creating environments, deploying application stacks, configuring networking—can and should be automated. EdgeForge AI is our way to do this reliably, on-premises, and without dependencies on American vendors.

It is also the first module of SentinelForge AI — the AIOps platform we are building to cover the entire lifecycle of open-source cloud infrastructure, from provisioning to predictive monitoring with CloudSentinel AI .

Want to see EdgeForge AI in action?

If you've been to WMF and want to see the full version, or if you haven't met us and want a live demo, please let us know.

Describe your infrastructure—what you manage, what the operational pain points are, how much time you spend on provisioning—and we'll set up a session where we use EdgeForge AI on your real-world environment.

→ Request a demo of EdgeForge AI