Deccan AI, a startup that helps frontier AI labs refine their models through post-training data and evaluation work, just closed a $25 million Series A round. The funding was led by A91 Partners, with Susquehanna International Group and Prosus Ventures also participating, as reported by TechCrunch AI.
The company, founded in October 2024, has built its business on a simple reality: even the most advanced AI labs need outside help making their models actually work in the real world.
What Deccan Actually Does
While companies like OpenAI, Anthropic, and Google DeepMind build core models in-house, the post-training work (generating expert feedback, running evaluations, building reinforcement learning environments) is increasingly outsourced. Deccan sits right in that gap.
Its services include:
- Helping models improve coding and agent capabilities
- Training systems to interact with external tools like APIs
- Running evaluation suites (its product Helix)
- Operations automation for enterprise clients
Google DeepMind and Snowflake are among its customers. The startup runs about two dozen active projects at any given time, according to founder Rukesh Reddy.
The India Bet
What makes Deccan stand out from competitors like Scale AI, Surge AI, Turing, and Mercor is its deliberate concentration in India. While rivals source contractors from 100+ countries, Deccan keeps most of its contributor base in one place, primarily operating out of Hyderabad alongside its San Francisco Bay Area headquarters.
“If you have operations in just one country, it becomes far easier to maintain quality,” Reddy told TechCrunch AI.
The numbers are significant: Deccan draws from a network of over 1 million contributors, with 5,000 to 10,000 active in a typical month. About 10% hold advanced degrees, and that share goes higher depending on project complexity. The company employs around 125 people full-time.
This approach puts a spotlight on India’s role in the global AI value chain, serving as a major supplier of training talent and data while frontier model development stays concentrated in the U.S. and China.
Why Post-Training Is Hard
Reddy made a point that matters for the whole industry: “Quality remains an unsolved problem.” Tolerance for errors in post-training is “close to zero” because mistakes directly affect how models perform in production.
This isn’t the same as basic data labeling. Post-training requires highly accurate, domain-specific data that’s genuinely difficult to scale. AI labs sometimes need large volumes of high-quality data within days, creating a constant tension between speed and accuracy.
The sector has also faced criticism over working conditions and pay for gig workers generating training data. Reddy said earnings on Deccan’s platform range from $10 to $700 per hour, with top contributors earning up to $7,000 monthly.
The Bigger Picture
Deccan describes itself as a “born GenAI” company, meaning it skipped the traditional computer vision labeling era and went straight to higher-skill work. The company grew 10x over the past year and now runs at a double-digit million-dollar revenue rate, though Reddy declined to share specifics.
One number worth watching: about 80% of revenue comes from its top five customers. That’s both a sign of how concentrated the frontier AI market is and a potential risk if any major client shifts strategy.
As AI models evolve beyond text into robotics and vision systems (what the industry calls “world models”), the demand for specialized post-training work is only going to grow. Deccan has also started sourcing talent from other markets, including the U.S., for niche expertise in areas like geospatial data and semiconductor design.
This $25M round positions Deccan to scale during a period when every major AI lab is racing to make their models more reliable. The full details are available in the original report on TechCrunch AI.