From Vibe Coding to Scalable Systems: Navigating the Era of AI-Assisted Engineering

Editor’s Note: At thegeekconf 2025, Arsh Goyal, AI enthusiast and engineer, met with Rakesh Ningthoujam, Head of Growth Marketing at GeekyAnts, to discuss the reality of vibe coding—the practice of using AI to rapidly build prototypes— and its impact on software development. While AI tools have made rapid prototyping more accessible, Arsh Goyal argues that scaling a product for the global stage still requires solid engineering skills. From the AI tools he relies on to his advice for aspiring developers, Arsh offers practical guidance for navigating an industry transformed by AI.

RN (Rakesh Ningthoujam): As an influencer and an engineer, you are likely reviewing new technology constantly. Which specific AI capability or tool are you personally using the most right now?

Arsh Goyal (AG): There are so many use cases now for writing, reviewing, and migrating code. I explore almost everything that comes out to see how it improves on the previous version. For core development, I use Cursor and Claude extensively across different use cases. I also look at how companies build MCPs (Model Context Protocols) and how to integrate them into various projects. For testing, I use TestSprite to automate test cases. On the content side, Gamma is great for presentations, and Napkin helps create the illustrations I need.

RN: What about "vibe coding" platforms—tools where you can just put an idea in front of people quickly?

AG: Those vibe coding tools like Lovable or Replit are very good for quick POCs (Proof of Concepts). If you have an idea and want to build a beta version to float among 50 or 100 users for feedback, it’s a great phase.

RN: Once you’ve "vibe coded" a POC, how do you move that to something scalable that works in production?

AG: Solid engineering is still required to solve for scale. AI has become much better at generating UIs—I don't have to write nearly as much HTML or CSS myself anymore—but you can't trust these automated tools with a massive user base yet. Tools like Riff (formerly Databutton) can integrate your database and handle backend tasks, and they are great for a pitch, but you still need a solid engineering team to scale to thousands of users.

RN: What percentage of that initial AI-generated code actually survives in the final production-grade product?

AG: It varies by project. For a major B2B use case, it might be less than 20%. For others, it could be 30% to 40%. You can tweak the frontend—the images and the content—but for the global market, you absolutely still need proper systems architects.

RN: Does that mean the "one-person company" built purely on vibe coding is a myth?

AG: It has been happening, and some are doing good revenue, but those founders usually have an engineering background. They aren't just vibe coding blindly; they either have that knowledge themselves or are getting deep technical consultation. If you don't understand the "in and out" of scaling, you risk security vulnerabilities and losing customer data. Craftsmanship is still vital.

RN: For the new generation of developers who have free access to Gemini, Perplexity, and ChatGPT, what should their approach be to stay relevant?