Shiv Shankar Dev

Vedic AI ↗

People ask deeply personal questions about career, relationships, family, and timing. Most online experiences either offer generic horoscopes or make the person do all the work of understanding a complicated chart.

Vedic AI approaches it differently. The app calculates the Kundali, uses relevant Jyotish information, and explains a traditional interpretation in language someone can follow.

Inside it, I've been building Yukti Drishti, an agent that can choose calculation tools, retrieve classical references, and connect evidence across a conversation. Getting that process to work reliably is far more involved than connecting a model to a chat box.

I lead product and engineering. My co-founder, Nisha Kumari, leads growth, customer insight and community. We're focused on making the core experience more useful and reliable.

When compact tool responses became enormous model inputs

While building Drishti, I found that some small-looking tool outputs were becoming large model inputs. Chart facts were being repeated across nested calls, even though the model didn't need to see them over and over.

We traced where the duplication was introduced, narrowed the information passed between steps, and kept the fuller records separately for debugging.

The most useful question turned out to be not how much evidence the system had, but which evidence the model needed at that moment.

Read the engineering note →

I also experiment with smaller language models and other approaches to computation. Training, evaluating, and debugging a model with limited compute has helped me understand what sits underneath an API.

Before Vedic AI, I worked on SwaraConnects, an earlier student-community project. That taught me a different lesson: building a product is not the same as getting people to use it.