Deccan AI
Verified CompanyBuild super accurate AI — post-training data, RL environments and agents
Frontier AI labs need expert humans to grade, correct and stress-test their models. Deccan AI, founded in 2023, built a network of over a million contributors — students, domain experts and PhDs — largely in India, and sells that work to Google DeepMind and Snowflake.
About Deccan AI
Every impressive AI model has an unglamorous layer underneath it: humans who read the model's output and say whether it is right. That layer is where Deccan AI lives. The company was founded in 2023 by Rukesh Reddy. It is headquartered in the San Francisco Bay Area, with its main operations team in Hyderabad and roughly 125 employees — but the number that actually describes the business is the other one: a network of more than a million contributors, including students, domain experts and PhDs, who do the work of generating expert feedback, grading model responses and building the environments models learn in. What it sells sits at what the industry calls the post-training layer. For frontier labs, that means supervised fine-tuning and RLHF datasets, expert evaluations, and reinforcement learning environments — the simulated settings in which a model practises and is scored. For enterprises, it means an evaluation suite called Helix and an operations automation platform the company markets as EnterpriseOS. The data work spans several domains. Multi-modal covers audio, video, visual and document-intelligence datasets for mid and post-training. Coding uses private code repositories built to capture real-world engineering scenarios, deployable against SWE-Bench, Terminal Bench or a customer's own benchmark. Agentic covers tool use, browser use, computer use and multi-step execution trajectories — the kind of data you need if you want a model to actually operate software rather than describe it. Robotics is a fourth stream. The customer list is the credential. Deccan AI says its customers include Google DeepMind and Snowflake — which is an unusual pair of names for a three-year-old company, and a fair signal of what the frontier labs are willing to outsource. In March 2026 it raised a $25 million Series A led by A91 Partners, with Susquehanna participating and existing investor Prosus Ventures following on. The round put it in direct comparison with Mercor and Scale AI, the American companies doing similar work — with the difference that Deccan AI sources most of its expert workforce from India, which is both its cost advantage and, given how many graduates the country produces in exactly the fields these models need grading in, its structural one. It is worth being precise about what this company is and is not. It does not train frontier models. It supplies the human judgement those models are trained against — a business that barely existed five years ago and now sits underneath most of the AI industry. Sources: deccan.ai (product lines, data domains and positioning, accessed August 2026); funding, founder, headcount, contributor network and customer details as reported by TechCrunch, Inc42 and Indian Startup News in March 2026. Images: Deccan AI, reproduced from the company's own website for identification.
Our Services
Post-Training Data — SFT and RLHF
Supervised fine-tuning and reinforcement-learning-from-human-feedback datasets built by domain experts rather than crowd workers. This is the layer that decides whether a model is merely fluent or actually correct, and it is what frontier labs buy from Deccan AI.
Reinforcement Learning Environments
Simulated environments in which models practise and are scored — the training grounds for agentic behaviour. Building them well requires people who understand both the domain being simulated and how models fail in it.
Helix — Evaluation Suite
The company's evaluation product for enterprises: structured, expert-graded assessment of how a model performs on the tasks a specific business actually cares about, rather than on public leaderboards.
EnterpriseOS — Agent Operations
An operations automation platform for deploying and running agents inside enterprises, positioned around reliability and measurable value rather than demonstrations.
Domain Data — Coding, Agentic, Multi-Modal and Robotics
Private code repositories capturing real engineering scenarios and deployable against SWE-Bench or Terminal Bench; agentic trajectories covering tool, browser and computer use; audio, video, visual and document-intelligence datasets; and a robotics stream.
Expert Talent Network
More than a million contributors — students, domain specialists and PhDs — recruited and quality-managed to do the grading, correction and adversarial testing that model training depends on. The company markets this as a route for the top 1% of domain experts to work on frontier AI.
Key Highlights
Our Portfolio
Post-Training Datasets
RL Environments
Helix
EnterpriseOS
Our Clients
Certifications & Partnerships
Recent News & Announcements
Deccan AI raises a $25 million Series A led by A91 Partners, with Susquehanna and existing investor Prosus Ventures participating
TechCrunch reports the company's AI training push relying on an India-based expert workforce of more than a million contributors
Company confirms Google DeepMind and Snowflake among its customers as it expands into agentic, coding and robotics data streams
Latest Articles
Alana Evans Net Worth in 2026 – Income Sources, Career & Earnings
Amy Lee Net Worth 2026: Income Sources, Evanescence Career & Personal Life
Abz Love Net Worth 2026: Five Star’s Career, TV Comeback & Income Breakdown
