Deccan AI Raises $25M Series A to Improve AI Model Reliability Through Post-Training

Call it what it is. The smartest models in the world still trip over the last mile. Deccan AI just raised $25M in Series A, and that round isn’t about chasing intelligence, it’s about forcing it to behave. Rukesh Reddy didn’t build another data shop dressed up in GenAI clothing. Deccan AI was born in the post-training trenches, where models either get sharper or get exposed. Mountain View address, Hyderabad muscle, Bengaluru expanding, and a contributor network north of 1M experts who don’t just label data, they interrogate it. That distinction is the whole game.

A91 Partners led the round, stepping into AI for the first time like they’ve been waiting for something worth the swing. Susquehanna International Group came in with them, and Prosus Ventures doubled down. Smart money tends to travel in packs when the signal is clean.

Here’s the part most people skip. Deccan AI is pulling 10x revenue growth with a double-digit million run rate, driven by roughly 10 enterprise and frontier lab customers, including Google DeepMind and Snowflake, plus a quiet majority of the so-called Magnificent 7. 80% of revenue tied to the top 5 customers. Concentration like that doesn’t happen unless you’re doing something mission critical, not optional.

And what they’re doing is building the layer nobody brags about but everyone bleeds over. STARK creates reinforcement learning environments that mirror real systems, not toy problems. Helix Evals blends human judgment with automated checks so models don’t hallucinate their way into production outages. EnterpriseOS Agents pushes AI into actual workflows where mistakes cost money, not just tweets. This is less about intelligence and more about consequences.

The network behind it matters just as much. 5,000–10,000 contributors active monthly, many with advanced degrees, earning anywhere from $10–$700 an hour. That’s not gig work, that’s precision labor. When your output demands near-zero error tolerance, you don’t crowdsource, you curate.

The takeaway is simple, but not easy. Everyone wants smarter models. Few are willing to invest in making them reliable. Deccan AI leaned into that gap early, built for it from day one, and now the market is catching up to their thesis.

Congratulations to Rukesh Reddy and the Deccan AI team, and credit to A91 Partners, Susquehanna International Group, and Prosus Ventures for recognizing that accuracy is the new scale. In a world sprinting toward more AI, the winners might just be the ones slowing it down long enough to make it right.