Mercor competitor Deccan AI raises $25M, sources experts from India

As request grows for grooming and refining AI models, Deccan AI — a startup supplying post-training information and valuation enactment — has raised $25 cardinal successful its archetypal large backing round, with overmuch of that enactment carried retired by an India-based workforce of experts.

The all-equity Series A circular was led by A91 Partners, with information from Susquehanna International Group and Prosus Ventures.

While frontier AI labs including OpenAI and Anthropic physique halfway models in-house, overmuch of the post-training enactment — from information procreation to valuation and reinforcement learning — is progressively being outsourced arsenic companies propulsion to marque systems reliable successful real-world use. Deccan is emerging arsenic 1 of a caller acceptable of startups serving that demand.

Founded successful October 2024, Deccan provides services ranging from helping models amended coding and cause capabilities to grooming systems to interact with outer tools specified arsenic exertion programming interfaces (APIs), which link AI models to bundle systems.

The startup works with frontier labs connected tasks specified arsenic generating adept feedback, moving evaluations and gathering reinforcement learning environments, portion besides serving enterprises done products including its valuation suite, Helix, and an operations automation platform. The enactment is besides evolving arsenic models determination beyond substance into alleged “world models” that amended recognize carnal environments, including robotics and imaginativeness systems.

Deccan’s customers see Google DeepMind and Snowflake, according to the company. It has onboarded astir 10 customers and runs a mates of twelve progressive projects astatine immoderate fixed time, laminitis Rukesh Reddy (pictured above) said successful an interview.

The startup, headquartered successful the San Francisco Bay Area with a ample operations squad successful Hyderabad, employs astir 125 radical and relies connected a web of much than 1 cardinal contributors, including students, domain experts, and PhDs. Around 5,000 to 10,000 contributors are progressive successful a emblematic month, Reddy told TechCrunch.

About 10% of Deccan’s contributor basal has precocious degrees specified arsenic master’s and PhDs, though the stock is higher among progressive contributors depending connected task requirements, Reddy said.

The marketplace for AI grooming services has expanded rapidly alongside the emergence of ample connection models, with companies specified arsenic Meta-owned Scale AI and its rival Surge AI, arsenic good arsenic startups Turing and Mercor competing to supply information labeling, evaluation, and reinforcement learning services.

“Quality remains an unsolved problem,” Reddy said, adding that tolerance for errors successful post-training is “close to zero” arsenic mistakes tin straight impact exemplary show successful production. That makes post-training much analyzable than earlier stages, requiring highly accurate, domain-specific information that is harder to scale.

The enactment is besides highly time-sensitive, helium said, with AI labs sometimes requiring ample volumes of high-quality information wrong days, making it hard to equilibrium velocity with accuracy.

The assemblage has faced disapproval implicit moving conditions and pay, with ample pools of gig workers often utilized to make grooming data. Reddy said net connected Deccan’s level scope from astir $10 to $700 per hour, with apical contributors earning up to $7,000 a month.

India emerges arsenic a hub for AI grooming talent

Even arsenic its customers are mostly U.S.-based AI labs, astir of Deccan’s contributors are based successful India. Competitors specified arsenic Turing and Mercor also root contractors from the country, but run crossed a broader acceptable of emerging markets.

Deccan chose to ore overmuch of its workforce successful India to amended negociate quality, Reddy said. “Many of our competitors spell to 100-plus countries to find the experts,” helium said. “If you person operations successful conscionable 1 country, it becomes acold easier to support quality.”

That attack highlights India’s existent presumption successful the planetary AI worth concatenation — arsenic a supplier of endowment and grooming information alternatively than a developer of frontier models, which stay concentrated among a fistful of U.S. companies and a fewer players successful China.

However, Reddy said Deccan has begun sourcing endowment from a fewer different markets, including the U.S., for niche expertise successful geospatial information and semiconductor design.

Reddy said Deccan was built arsenic a “born GenAI” company, successful opposition to accepted information labeling firms that began with machine imaginativeness tasks. This means it has focused connected higher-skill enactment from the outset.

Deccan grew 10x implicit the past twelvemonth and is present astatine a double-digit million-dollar gross tally rate, Reddy said, declining to stock specifics. About 80% of its gross comes from its apical 5 customers, reflecting the concentrated quality of the frontier AI market, helium added.