AI impresses Xhoni Shollaj almost every day, from protein folding and the idea of a virtual cell to autonomous driving and robotics. In this conversation, host Daniel Kazani follows Xhoni’s journey from business studies and data roles at PwC and EY in Albania and Bulgaria to a Master of Science at the National University of Singapore and his current work as a senior AI solutions architect at NVIDIA.

ㅤ

The discussion moves from early natural language processing and computer vision projects, document reduction and summarization tools, to building and maintaining large scale language model applications. Xhoni shares how staying in touch with GitHub trending projects, arXiv style paper feeds, and the open source community shaped his path. Founders, CTOs and decision makers hear concrete talk on AI experiments versus production systems, scalability, security, hallucinations, golden datasets, vibe coding, tools like Cursor, ChatGPT and Gemini, and why contributing to open source with teams at NVIDIA, Google and others can be a powerful way to stand out.

ㅤ

👤 Guest Bio

Xhoni Shollaj is a Senior AI Solutions Engineer at NVIDIA, specializing in developing and deploying large language model architectures. He started with business, moved into computer science, and began his AI journey in research and development teams at PwC and EY in Albania and Bulgaria, building machine learning based applications and automation solutions. Xhoni then joined the National University of Singapore, working on internal automation tools and research support for patents and papers, before moving into his current role at NVIDIA in Asia.

ㅤ

📌 What We Cover

• How Xhoni moved from business studies and data roles at PwC and EY in the Balkans to a Master of Science at the National University of Singapore and into AI solutions work at NVIDIA.
• Why he chose Singapore for its faculty, research direction and blend of cultures, and how being location agnostic helped him follow the strongest data science programs.
• The habits he sees as most useful for people who want to succeed in AI, including staying in touch with the latest technologies, GitHub trending, arxiv style feeds, open source projects and strong news sources.
• Areas where AI feels most disruptive today, from protein folding and the path toward a virtual cell for drug discovery and disease treatment to space exploration, SpaceX and ideas like space data centers.
• How to distinguish an AI experiment or small POC from a production system, with concrete points on autoscaling, multi cloud and multi zone backups, security pipelines, identity and access management, encryption, multilingual behavior, hallucination tracking and observability.
• Approaches to accuracy and hallucinations, including well built RAG pipelines, choosing the right benchmarks and metrics, literature reviews, leaderboards, human in the loop evaluation and tracing problems back to data sources or model behavior.
• The reality of vibe coding for non technical founders, why it is a net positive and equalizer, and how to combine fast POCs with later help from experienced engineers on scaling, security and edge scenarios.
• Tools and workflows Xhoni personally uses, such as Cursor with Claude 4.5 Sonnet, ChatGPT and Gemini for brainstorming, creating plans, testing ideas and even asking models to make fun of an idea to expose weak points.
• The most common challenge companies face when integrating AI into their business, why a golden dataset and clean, validated, well reviewed data can make or break a project, and how synthetic data and diverse scenarios help test chat bot performance.
• Why AI systems in sectors like hospitality need clean booking and address data, strong formatting, and synthetic test scenarios with mixed languages, toxic and non toxic inputs, and special characters to evaluate real world behavior.
• Thoughts on interpretability and mechanistic interpretability, the black box nature of transformer layers today, and why being able to trace reasoning in sensitive areas like healthcare or drug simulation matters.
• How different large models like OpenAI, Claude, Gemini, DeepSeek and NVIDIA models feel to users because of data sources such as Reddit and prompt level instructions, leading to different levels of confidence, politeness and directness.
• What a billion plus dollar AI data center collaboration between NVIDIA and Deutsche Telekom in Germany might mean for telecom, communication, research and startup ecosystems in Munich, Germany and across Europe.
• Final advice for students, graduates and people struggling in the current environment, including a clear call to contribute in their free time to open source NVIDIA and Google projects, build relationships, learn in public and stand out through real work.

ㅤ

🔗 Resources Mentioned

• NVIDIA
• PwC
• EY
• National University of Singapore
• AWS
• Azure
• Deutsche Telekom
• SpaceX
• GitHub trending
• Arxiv style feeds and “alpha Arxiv sanity”
• Cursor
• Claude 4.5 Sonnet
• ChatGPT
• Gemini
• DeepSeek
• Tinker
• OpenAI
• Anthropic
• Google
• Reddit