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

As demand grows for coaching and refining AI fashions, Deccan AI — a startup supplying post-training knowledge and analysis work — has raised $25 million in its first main funding spherical, with a lot of that work carried out by an India-based workforce of consultants.

The all-equity Collection A spherical was led by A91 Companions, with participation from Susquehanna Worldwide Group and Prosus Ventures.

Whereas frontier AI labs together with OpenAI and Anthropic construct core fashions in-house, a lot of the post-training work — from knowledge technology to analysis and reinforcement studying — is more and more being outsourced as firms push to make programs dependable in real-world use. Deccan is rising as one among a brand new set of startups serving that demand.

Based in October 2024, Deccan gives companies starting from serving to fashions enhance coding and agent capabilities to coaching programs to work together with exterior instruments reminiscent of utility programming interfaces (APIs), which join AI fashions to software program programs.

The startup works with frontier labs on duties reminiscent of producing skilled suggestions, working evaluations and constructing reinforcement studying environments, whereas additionally serving enterprises by means of merchandise together with its analysis suite, Helix, and an operations automation platform. The work can be evolving as fashions transfer past textual content into so-called “world fashions” that higher perceive bodily environments, together with robotics and imaginative and prescient programs.

Deccan’s clients embody Google DeepMind and Snowflake, in line with the corporate. It has onboarded about 10 clients and runs a few dozen lively tasks at any given time, founder Rukesh Reddy (pictured above) stated in an interview.

The startup, headquartered within the San Francisco Bay Space with a big operations staff in Hyderabad, employs about 125 folks and depends on a community of greater than 1 million contributors, together with college students, area consultants, and PhDs. Round 5,000 to 10,000 contributors are lively in a typical month, Reddy informed TechCrunch.

About 10% of Deccan’s contributor base has superior levels reminiscent of grasp’s and PhDs, although the share is larger amongst lively contributors relying on undertaking necessities, Reddy stated.

The marketplace for AI coaching companies has expanded quickly alongside the rise of huge language fashions, with firms reminiscent of Meta-owned Scale AI and its rival Surge AI, in addition to startups Turing and Mercor competing to offer knowledge labeling, analysis, and reinforcement studying companies.

“High quality stays an unsolved downside,” Reddy stated, including that tolerance for errors in post-training is “near zero” as errors can straight have an effect on mannequin efficiency in manufacturing. That makes post-training extra complicated than earlier phases, requiring extremely correct, domain-specific knowledge that’s tougher to scale.

The work can be extremely time-sensitive, he stated, with AI labs generally requiring giant volumes of high-quality knowledge inside days, making it tough to stability pace with accuracy.

The sector has confronted criticism over working circumstances and pay, with giant swimming pools of gig staff typically used to generate coaching knowledge. Reddy stated earnings on Deccan’s platform vary from about $10 to $700 per hour, with high contributors incomes as much as $7,000 a month.

India emerges as a hub for AI coaching expertise

Whilst its clients are largely U.S.-based AI labs, most of Deccan’s contributors are based mostly in India. Opponents reminiscent of Turing and Mercor additionally supply contractors from the nation, however function throughout a broader set of rising markets.

Deccan selected to pay attention a lot of its workforce in India to raised handle high quality, Reddy stated. “A lot of our opponents go to 100-plus international locations to seek out the consultants,” he stated. “In case you have operations in only one nation, it turns into far simpler to keep up high quality.”

That strategy highlights India’s present place within the international AI worth chain — as a provider of expertise and coaching knowledge relatively than a developer of frontier fashions, which stay concentrated amongst a handful of U.S. firms and some gamers in China.

Nonetheless, Reddy stated Deccan has begun sourcing expertise from a couple of different markets, together with the U.S., for area of interest experience in geospatial knowledge and semiconductor design.

Reddy stated Deccan was constructed as a “born GenAI” firm, in distinction to conventional knowledge labeling corporations that started with laptop imaginative and prescient duties. This implies it has centered on higher-skill work from the outset.

Deccan grew 10x over the previous 12 months and is now at a double-digit million-dollar income run charge, Reddy stated, declining to share specifics. About 80% of its income comes from its high 5 clients, reflecting the concentrated nature of the frontier AI market, he added.