In the software world, we usually lose sleep over bugs in our own code. But in late 2023, while I was serving as Senior Manager at PETRONAS, I hit a wall that had nothing to do with my code—and everything to do with dependency.

We were days away from a major production launch of J.AI. The system was tested, the stakeholders were ready, and the hype was massive. Then, the lights flickered.

​Two things happened simultaneously:

1. ​The Global GPU Crunch: Suddenly, getting compute capacity became nearly impossible.
2. ​The Token Limit: OpenAI, struggling with massive global demand, imposed strict rate limits.

​Overnight, our application’s “brain” simply stopped responding. As a Senior Manager responsible for delivery, I watched us get locked out of our own product because we relied 100% on a single vendor’s API. We were beholden to a service provider that didn’t know we existed, and we were powerless to fix it.

​The “Not-so-Beautiful” Fix

​Most teams in that situation would have delayed the launch. We couldn’t.

​So, I stopped acting like a manager and started acting like a plumber. I had to do something very unusual for that time: I started scavenging.

We found chunks of compute capacity in different testing environments. We wrote scripts to break our big requests into smaller pieces. If one environment was busy, we manually rerouted the data to another. We literally patched together inference capacity from scraps just to keep the lights on.

​It was a “Frankenstein” solution, messy and complex but it worked. We survived.

​The Business Lesson: Don’t Rent Your Foundation

That sleepless week was a wake-up call. I realized that by hardcoding our app to a specific model provider (like import openai), we had given them total control over our destiny.

If a giant like PETRONAS could be crippled by a single vendor’s API limit, what chance does a normal business have?

​I realized that enterprises needed a way to decouple their work (the business logic) from the worker (the AI model). You shouldn’t have to rewrite your software just because a model provider changed their pricing, hit a capacity limit, or discontinued a model.

​That is why I built Skhy Vidur.

What is Skhy Vidur?

​Think of Skhy Vidur as a Universal Power Adapter for Artificial Intelligence.

When you travel, you don’t rewire your laptop for every country; you just use an adapter. Skhy Vidur is that adapter for your business. It sits between your company and the chaotic world of AI models.

Here is how it solves the problem I faced in 2023:

​1. The Manager (The Switchboard)

Instead of your app talking directly to Google or OpenAI, it talks to the Vidur Manager.

• ​The Scenario: You usually use GPT-5 or Gemini. Suddenly, it goes down or gets too expensive.
• With Vidur: You just change a setting in the dashboard to route tasks to a cheaper model (like Llama 3 or Qwen or Deepseek) running on your own servers or any other sovereign GPU provider. Your app doesn’t know the difference; it just keeps working.

​2. Datac (The Quality Control)

During that crisis, we realized that bad data wastes precious processing power. Datac is a hygiene layer. It grabs data (from the web or files), cleans it, and removes the “garbage” before you pay to process it. It ensures you only spend money analyzing high-quality information.

​3. The Client (The Secure Room)

For your most sensitive data—financials, legal strategy, defense logs—you can’t send that to the cloud at all. The Vidur Client runs on your own hardware, completely offline. It does the thinking inside your four walls.

​The Future is Sovereign

​We are entering the era of Physical and AI Sovereignty.

​The businesses that win in the next decade won’t be the ones with the best prompts. They will be the ones who own their infrastructure. They will be the ones who can switch from a $20/million-token model to a free local model with the click of a button.

We built Skhy Vidur so that no CTO or Founder ever has to feel the panic I felt in 2023.

Stop building wrappers. Start building architecture.