In 2024, the world was deep in the AI wave. I was sitting in Dhangadhi, Nepal watching the entire landscape shift. I had been running XyleHosting for a couple of years at that point, but something about the AI wave felt different — it felt like the kind of shift you either get on top of or get left behind by.
I decided to get on top of it.
The Idea
I wanted to build an AI-powered search engine — something that could give better, more contextual answers than a traditional keyword search. The core idea was simple: combine a fast search index with a language model to deliver results that actually make sense, not just a list of blue links.
"I had zero experience in machine learning. That stopped me for about 48 hours."
I spent those 48 hours reading documentation, watching tutorials, and testing APIs. I quickly realized that building from scratch would take years, but integrating existing AI APIs with a solid search architecture was achievable.
The Build
CoreX AI was built in stages. The first version was barely a prototype — it took a query, hit a few data sources, ran the result through an AI model, and returned a synthesized answer. The latency was terrible and the answers were hit or miss, but it worked.
The key challenges we worked through:
- Speed — Users expect search results in under a second. Getting AI inference that fast required careful optimization and caching strategies.
- Accuracy — Raw LLM output can confidently say wrong things. We built verification layers and source attribution to keep answers grounded.
- Scaling — Going from 100 queries a day to 100,000 required a complete rethinking of the infrastructure.
- Cost — AI API calls are not cheap. We had to get very creative about when and how we call models versus serving cached results.
Launch and Growth
We launched CoreX AI quietly, sharing it in tech communities and forums. The response was positive immediately. People loved that it gave direct answers rather than just links. Within weeks we had thousands of users, and within a few months the platform was handling tens of thousands of daily queries.
CoreX AI grew to over 97,000 registered users and processed more than 100,000 searches every single day at its peak — growth that happened almost entirely through word of mouth. The project has since closed, but the lessons it taught me about building and scaling AI products have been invaluable.
Why CoreX AI Closed
CoreX AI has since closed. The honest reason: lack of team and time. Running an AI search engine at scale is operationally very demanding. AI API costs, infrastructure, model quality monitoring, content moderation — it all adds up, and it all requires dedicated people working on it full-time.
At the same time, XyleHosting was growing fast and needed my full attention. I was stretched between two demanding platforms, and I couldn't give either of them the focus they deserved. Rather than let CoreX AI stagnate or deliver a poor experience, I made the decision to shut it down properly.
No regrets. CoreX AI taught me an enormous amount about AI systems, product scaling, and infrastructure — lessons that directly benefit XyleHosting today.
What I Learned
Building an AI product is fundamentally different from building a traditional software product. The behavior is non-deterministic — the same input can produce different outputs. This means your testing and quality assurance process has to evolve. You can't just write unit tests and call it done.
I also learned that AI products live and die on trust. If users get a confident but wrong answer twice, they leave and don't come back. Building systems that are transparent about uncertainty — and that clearly cite sources — was one of the most important decisions we made.
And perhaps most importantly: you don't need to have a PhD in machine learning to build AI products. What you need is curiosity, persistence, and a willingness to learn in public.
