The Banker Who Built An AI Factory

On This Podcast

In this conversation, we examine how a chance conversation with a stranger on a flight introduced Nick Gushin to ChatGPT's ability to write code, sparking his self-directed journey into Python. Starting from a simple print hello world command, he taught himself programming using ChatGPT and YouTube podcasts as his tutors, gradually learning to deploy applications and work with cloud infrastructure. Nick describes a nine-month stretch of coding twelve to fourteen hours a day, treating the process as an intensive full-time education even while still formally employed in banking. He built more than twenty pet projects in Python, most of which failed to find an audience, but one Telegram chatbot grew to over a hundred thousand users, giving him confidence that his new skills could produce something real.

We also discuss why Nick stepped away from an eighteen-year banking career, despite valuing the industry's role in society, once he saw how much faster he could iterate on ideas with AI. He describes setting deliberate milestones during a sabbatical, first testing whether he could code independently and then whether his products could attract real users, before one chatbot project drew over a thousand active users within a month. Nick contrasts this develop-fast, fail-fast approach with banking, where bringing a single new product to market could take nine to twelve months, meaning a whole career might only produce a handful of launched ideas. He credits this rapid iteration cycle, learning new protocols and libraries with every project, for accelerating his growth far beyond what a corporate role could offer.

A third focus of the conversation is the architecture behind the Swiss AI Chatbot Factory, which automates an eight-step chatbot-building process that Nick and his co-founder Robert originally performed by hand for clients. We examine how user input is never passed directly to a response-generating function; instead it moves through layered AI agents that check relevance, screen for jailbreak attempts, and combine vector search with linguistic search against the client's data. Nick recounts an internal test showing that aggressive user language pushed models toward hallucinating answers more than ninety percent of the time, which shaped the strict verification rules now built into the pipeline. The result, he says, is a chatbot that can be generated from a company URL in under three minutes while still resisting the fabricated or off-topic answers that plague simpler chatbot implementations.