As demand for AI model training and refinement increases, Deccan AI, a startup providing post-training data and evaluation services, has secured $25 million in its initial major funding round, predominantly supported by an India-based expert workforce.
The all-equity Series A funding was led by A91 Partners, with involvement from Susquehanna International Group and Prosus Ventures.
While frontier AI laboratories such as OpenAI and Anthropic develop core models internally, much of the post-training processes—ranging from data generation to evaluation and reinforcement learning—are outsourced as firms aim to ensure system reliability in practical applications. Deccan is among the emerging startups addressing this market need.
Established in October 2024, Deccan offers services that enhance model capabilities in coding and agent interaction while training systems to engage with external tools like APIs, connecting AI models with software systems.
The startup collaborates with frontier labs on tasks such as generating expert feedback, conducting evaluations, and constructing reinforcement learning environments, alongside serving enterprises through products like its evaluation suite, Helix, and an operations automation platform. The work is advancing as models extend beyond text to encompass “world models” that comprehend physical environments, including robotics and vision systems.
Deccan’s clientele includes Google DeepMind and Snowflake. It has approximately 10 clients and manages several dozen ongoing projects, according to founder Rukesh Reddy.
Headquartered in the San Francisco Bay Area with a significant operations team in Hyderabad, Deccan employs about 125 individuals and leverages a network of over 1 million contributors, encompassing students, domain experts, and PhDs. Typically, 5,000 to 10,000 contributors are active monthly.
About 10% of Deccan’s contributor base possesses advanced degrees such as master’s and PhDs, with a higher proportion among active contributors depending on project needs.
The AI training services market has expanded rapidly alongside large language models’ emergence, with companies like Meta-owned Scale AI and its competitor Surge AI, as well as startups Turing and Mercor, competing to provide data labeling, evaluation, and reinforcement learning services.
“Quality remains an unresolved issue,” Reddy stated, emphasizing that error tolerance in post-training is “virtually nonexistent,” as inaccuracies can directly impact model performance in production. This complexity makes post-training more challenging than earlier stages, necessitating highly accurate, domain-specific data that is difficult to scale.
The work is also time-sensitive, with AI labs occasionally needing large volumes of high-quality data within days, complicating the balance of speed and accuracy.
The sector has faced scrutiny over working conditions and pay, with gig workers frequently used for training data generation. Reddy mentioned that earnings on Deccan’s platform range from approximately $10 to $700 per hour, with top contributors earning up to $7,000 monthly.
India has emerged as a prominent hub for AI training talent. Despite its predominantly U.S.-based clientele, most of Deccan’s contributors are located in India. Competitors such as Turing and Mercor also source contractors from India but operate across a broader range of emerging markets.
Deccan chose to concentrate much of its workforce in India to better manage quality. “Many competitors engage experts from over 100 countries,” Reddy said. “Operating within one country simplifies quality maintenance.”
This approach underscores India’s role in the global AI value chain as a provider of talent and training data, while frontier model development remains largely concentrated among a few U.S. and Chinese firms.
Nonetheless, Reddy noted Deccan is sourcing talent from other markets, including the U.S., for specialized expertise in areas like geospatial data and semiconductor design.
Reddy described Deccan as a “born GenAI” company, distinguishing it from traditional data labeling firms that initially focused on computer vision tasks, by concentrating on high-skill work from the start.
Deccan grew tenfold over the past year and has reached a double-digit million-dollar revenue run rate, though Reddy withheld specific figures. About 80% of its revenue is derived from its top five customers, a reflection of the concentrated nature of the frontier AI market.