Startups entering the generative AI space may consider fine-tuning models on proprietary data a necessary step. For Geektrust, a tech hiring platform with five years of labeled data from their earlier product, Codu.ai, this would appear to be a logical choice. Yet, they intentionally decided against it.

Through the lens of CEO Krishnan Nair, this article explores how Geektrust leverages prompt engineering, makes architectural choices to support long-form dialog, and prioritizes delivery over optimization.

Speed over perfection

Geektrust aimed to solve a pressing user problem early, without waiting for a perfect solution.

“It's not that we don't have data. We have the data, but optimization is not what I'm looking for. That 90-95% is good enough. First, I want to get you as a customer to prove it works.”

This principle has allowed the team to iterate faster, avoid technical bottlenecks, and maintain focus on business goals. For startups with limited resources, this model presents a pragmatic alternative to time and resource-intensive fine-tuning.

Even at the operational level, Krishnan embraces this principle by opting for manual processes where necessary to ensure the core AI solution can be deployed and validated quickly. He directs his team to "sell B2B first. Everything else you'll handle manually because you want to prove that customers will pay for it."

A Grandmother-Child framework for prompt engineering

Rather than adapting the model to the data, Geektrust adapted the prompts to the model. Prompt engineering serves as the primary mechanism for achieving relevance and precision in application output.

Krishnan compares their approach to communicating with a wide-ranging audience, coining a new Grandmother-Child two-hander:

"You have to explain things to OpenAI like you would to your 90-year-old grandmother. And when you want it to do things for you, you have to give instructions like you would do a five-year-old.”

Designing for context: Moving beyond RAG

Geektrust initially explored retrieval-augmented generation (RAG) to handle technical interviews. However, they encountered limitations in how RAG maintained context over multiple conversational turns.

"When we used RAG, there was a drop in context. In our use case, this is not data that is getting shared from a document – it's a 60 to 90 minute conversation.”

The retrieval mechanism, while effective for short tasks, struggled to track evolving interactions.

Instead, the team developed an agentic workflow that uses dynamic context injection. The core of their solution involves continuously "feeding in business data" into the prompts during the hiring interview. After a candidate provides a response, the system analyzes it, determines what the interviewer cares about for the next question, and injects that relevant information back into the prompt. This iterative process ensures that subsequent questions are relevant and build upon the ongoing dialog, mimicking a human interviewer's ability to adapt and follow up.

Final thoughts

Simplicity is a deliberate and effective choice. Rather than investing early in model customization, Geektrust optimized for clarity, responsiveness, and speed.

For leaders considering their own roadmap: solving real problems with well-structured inputs can sometimes achieve more than technical refinement. Focus on what matters at the right stage of growth.