Rukesh Reddy, founder of Deccan AI, and Purva Bhandari sit down with Rohit Yadav to explain where the data layer sits in the model building stack, and why it does not disappear as models get smarter. Every model has three ingredients: compute, the engineers writing the algorithms, and the data. Deccan builds the third one, supplying training and eval data to the handful of labs building frontier models, and now also helping enterprises deploy AI.

Rukesh came to this from finance, working for a bank in New York before what he calls a complete U-turn. The thesis was a shift in what annotation means. The industry scaled around 2014 with self driving, when the work was bounding boxes, accurate but not IQ heavy. Post ChatGPT it needed engineers, PhDs and domain experts who can reason with a model, which he saw as a right to win for a company delivering out of India. The name is deliberate too: after learning that Cisco is a homage to San Francisco, he asked why nobody was claiming the Deccan, India’s own corridor across Hyderabad, Bangalore and Pune. The company bootstrapped for 18 months, raised a small round, then a Series A in March.

Whether you build models, deploy them inside an enterprise, or invest in either, this is a clear look at the layer most people underestimate.

🔑 Key insights you will walk away with:

➡️ Why the data layer survives every change in algorithm, from RLHF through to RL

➡️ The difference between full stack and first party models, and why those lines keep blurring

➡️ Why in house data teams at labs are a shrinking share of the market

➡️ How synthetic and human data fused into a hybrid, and why one or two percent of human input is the kicker

➡️ Why knowing where models break is harder than generating data, and what that demands of a platform

➡️ The case for starting in one country when cycle times can be as short as a day

➡️ Why today’s models have high bookish knowledge and struggle with the exceptions outside codified process

➡️ Why the binding constraint on open source adoption is engineering capacity rather than licensing

➡️ Why ARR is a shaky measure when revenue is far more volatile than previous tech waves

➡️ Why language is a flattened out view of the world, and what world models are actually for

Links:

Rukesh Reddy: https://www.linkedin.com/in/rukeshreddy/

Purva Bhandari: https://www.linkedin.com/in/purvabhandari/

Deccan AI: https://www.deccan.ai/

Rohit Yadav: https://www.linkedin.com/in/rohityadav23/

Newsletter: https://yadavrohit.substack.com/