The Startup That Raised $25 Million to Put Humans Back Into AI

For much of the generative AI boom, the industry’s attention was fixed on two inputs: compute and talent. 

NVIDIA became one of the world’s most valuable companies, with a $5 trillion market cap, as hyperscalers committed billions of dollars to data centres and frontier labs competed aggressively for top researchers.

But training a frontier model is only a part of the challenge. Teaching the models to follow instructions, reason reliably, use tools, and operate inside real-world workflows is something critical. 

That work increasingly depends on post-training data, evaluation systems, reinforcement-learning environments and human feedback.

It is in that layer of the AI stack that Deccan AI, an AI evaluation startup, has built its business. Headquartered in the United States with most of its operations in India, the company says it grew tenfold in a year while supplying post-training data to a majority of the Magnificent Seven.

The company builds post-training datasets and reinforcement learning environments where models cycle through trial and error, covering coding, multimodal, research, safety, and domain-specific data. The company also tests frontier models and agents for accuracy, reasoning, safety, and reliability through both automated scoring and human expert review. 

“Post-training is where frontier model makers need to ensure that the model is able to follow instructions properly, and get the right domain expertise. This requires experts to give feedback,” says Sajo Mathews, who leads machine learning efforts at Deccan AI, in an interaction with AIM. 

Internally, Deccan AI packages these capabilities into products such as STARK, its reinforcement-learning environment platform, and Helix, an evaluation system for benchmarking and testing AI agents.

According to the company, it has an expert network of over a million registered professionals, with more than 12,000 vetted specialists available for complex projects. The domains covered span software engineering, medicine, finance, law, psychology, research, design, linguistics, and consulting

“There are things that the frontier AI models are still not good at, and we need humans to come in and teach these models to become better,” he says. 

These experts go beyond basic labelling and review model outputs, rank AI responses, create benchmark tasks, evaluate agents, and provide domain-specific judgments that automated scoring cannot reliably produce. 

“Which is why we have experts who have PhDs, and doctors as a part of the talent pool,” says Mathews. They also help create evaluation benchmarks and reasoning tasks that are later used to measure model performance during deployment.

Increasingly, this feedback is also being used to train agentic systems that interact with environments rather than merely predict text, an area where Deccan says reinforcement-learning infrastructure is becoming as important as traditional dataset creation.

That market is growing rapidly. A recent independent report by Epoch AI found that contracts for reinforcement-learning environments often run into six- and seven-figure quarterly engagements, with frontier labs increasingly investing in simulated environments where models can learn through trial and error. 

The report noted that complex software-engineering tasks can occasionally cost as much as $20,000 each to create, while individual environment contracts can reach well into the millions of dollars.

In March, Deccan AI raised a $25M Series A round and is backed by investors including A91 Partners, Prosus Ventures, and Susquehanna. While the company hasn’t officially listed its largest customers, it counts Google DeepMind and Snowflake among its customers.

While Deccan AI today operates in a competitive field that includes companies like Scale AI, Mercor and Turing, the company says these players run highly distributed pipelines in which annotators, project managers and quality reviewers operate in silos. 

The Post Training Stack

This is where the company wants to make a difference, by positioning itself as an operations-intensive alternative. Deccan AI says that annotators are trained in person, not via video recordings, and that product and client-facing teams sit in the same time zone and often in the same building. 

Unlike firms that focus primarily on sourcing talent or collecting annotations, Deccan is attempting to build a full post-training stack spanning training environments, evaluation systems, expert feedback loops and enterprise deployment.

When clients switch from larger competitors, the company says they notice that the feedback loop between what a client asks for and what an annotator produces remains intact, rather than getting lost across layers.

In addition to these two layers, the company is actively providing enterprise AI solutions by building and deploying agents for business workflows, with humans in the loop for oversight. 

“We see a big gap in how enterprises are actually deploying AI models in a reliable and safe manner,” says Mathews. “You broadly need two things to do so—a system and an architecture which amplifies the good parts of the AI model and mitigates the others.” 

That architecture increasingly combines automated evaluations with expert intervention, creating a feedback loop in which deployment data can inform future model improvements.

The Enterprise AI Moat

The company refers to this deployment layer as EnterpriseOS, a platform designed to run AI agents inside enterprise workflows while routing sensitive or uncertain decisions through human reviewers.

The focus on deployment comes at a time when even frontier model providers are acknowledging that better models alone do not guarantee enterprise adoption. 

In May, OpenAI launched the OpenAI Deployment Company, a dedicated organisation built around helping customers integrate AI into real-world workflows through teams of forward-deployed engineers. 

Anthropic, too, has been expanding its Applied AI and deployment-focused teams, reflecting a broader recognition among frontier AI labs that getting models into production is often a harder problem than improving benchmark scores. 

Even at this enterprise AI layer, Deccan AI’s assets are the same human experts. Mathews says their expert networks help evaluate the model's sensitive parts and provide feedback. At the other end of the spectrum, these experts can take over the entire service and run it for the organisation. 

In effect, the same expert network that helps train and evaluate models can also serve as an operational fallback when automation is not yet reliable enough for fully autonomous deployment.

Deccan AI today has a 400+-person office in Hyderabad and a new hub in Bangalore, which is set to launch soon, and is now targeting the Global Capability Centres (GCCs) market in India. These centres handle back-office and operational functions for major multinationals. 

“A lot of the real operational knowledge of how those operations are done today probably exists more in India than it does even with the headquarters. You cannot do it in a vacuum and disconnected from reality,” he says, indicating the opportunity that exists with DeccanAI today to serve these GCCs. 

Deccan AI has also recently joined NASSCOM as it formalises GCC partnerships. The company says its goal is to help GCCs transition into AI-native operations through data products, human-in-the-loop services, or full managed service arrangements.