Ayush Bhardwaj, drawing from his experience in both hedge funds and a pharma tech startup, presents insights into building applied vertical AI—AI tailored for specific industries. He defines applied vertical AI as AI designed to simulate the job of a person within a particular sector, contrasting it with general-purpose AI like Google Translate. Despite expectations that transitioning from the fast-paced hedge fund environment to the slower, meticulous pharma industry would drastically change his work, Bhardwaj found the core applied AI processes remarkably similar across these verticals. He also notes that many industries, including legal and real estate tech, follow comparable approaches in developing vertical AI solutions.
Bhardwaj outlines a seven-step recipe for building applied vertical AI. The first step is to formulate a narrowly defined problem rather than attempting to solve everything at once. Next, identifying and curating proprietary data is crucial, as this data differentiates the AI’s performance from general models like ChatGPT. The third step involves modeling the problem through carefully crafted prompts that mimic how a human expert would approach the task. Observability follows, enabling developers to monitor and debug the AI’s behavior effectively. These initial steps are relatively straightforward and can be implemented quickly by skilled engineers.
The real challenge, Bhardwaj explains, lies in iteration and evaluation. Unlike coding or math tasks where outputs can be objectively verified, vertical AI applications in finance or pharma lack clear-cut answer keys, making it difficult for engineers to judge model performance without domain expertise. He emphasizes that large language models (LLMs) cannot self-verify meaningfully in these contexts, as they generate plausible text without true understanding. Moreover, proprietary data in these fields is often scarce or hidden due to competitive reasons, further complicating model training and validation.
To overcome these challenges, Bhardwaj stresses the importance of hiring domain experts—the actual users of the AI tools—to guide development. These experts help refine prompts, curate data, and create logical workflows that reflect real-world problem-solving. By embedding these users in a continuous learning loop, the AI system gradually improves through supervised fine-tuning, reinforcement learning from human feedback, rubric-based evaluation, and error analysis. This iterative process generates valuable proprietary data and ensures the AI delivers tangible value, such as saving time or generating alpha in finance or pharma.
In conclusion, Bhardwaj highlights that while AI infrastructure and models are widely accessible commodities, the true competitive advantage lies in domain expertise and proprietary data. Applied vertical AI must justify its return on investment immediately, especially in high-stakes industries like finance and pharma, where failure to deliver results leads to swift rejection. Human experts remain central to the loop, with AI serving as an assistant rather than a replacement. The path to advanced AI capable of causal reasoning and autonomous decision-making—what Bhardwaj calls the “AGI line”—is still distant, making domain knowledge and data the enduring moats for successful vertical AI ventures.
