Deccan AI raises $25 million to scale post-training work

Deccan AI has raised $25 million in a Series A round led by A91 Partners with participation from Susquehanna International Group and Prosus Ventures, according to TechCrunch. The startup, founded in October 2024, sells “post-training” work—expert feedback, evaluations and reinforcement-learning environments—to frontier model builders and enterprises, and says customers include Google DeepMind and Snowflake. Headquartered in the San Francisco Bay Area, Deccan runs a large operations team in Hyderabad and relies on a contributor network it puts at more than one million people.

The pitch is less about building the next flagship model than about making existing models behave in production. TechCrunch quotes founder Rukesh Reddy describing “quality” as the unsolved problem, with “close to zero” tolerance for errors once models are deployed inside products. That framing points to a quiet shift in the AI supply chain: as base models become commoditised and widely licensed, the scarce input is increasingly the labor needed to test, correct, and harden them—often under tight deadlines. Deccan says it typically has a few dozen active projects and needs high-quality outputs “within days,” a tempo that pushes work toward large pools of on-call contractors.

The company’s workforce model also shows why “AI automation” keeps creating new job categories rather than simply removing old ones. Post-training work is not generic data labelling; it can require domain expertise, careful scoring, and consistent adherence to rubrics that change as models and benchmarks change. Deccan says around 10% of its contributor base has advanced degrees, and that the share is higher among active contributors depending on the project. Yet the bargaining power sits with the labs that set acceptance criteria and can reject work at scale, while the cost of rework is pushed onto the contractor layer.

Deccan’s decision to concentrate much of its contractor supply in India—rather than sourcing across “100-plus countries,” as Reddy describes competitors—reads as an operational choice to control variance. A single jurisdiction can mean more predictable recruiting pipelines, more uniform training, and easier dispute resolution when quality fails. It also concentrates a strategic dependency: if post-training becomes the bottleneck for shipping reliable systems, then the place where that work is organised matters as much as where the model weights are trained.

TechCrunch reports Deccan claims earnings on its platform range from about $10 to $700 per hour, with top contributors earning up to $7,000 a month. The spread is the story: the same pipeline that can pay a niche specialist handsomely can also sustain a large base of low-paid, time-sensitive piecework, all wrapped in the marketing language of “AI.”

Deccan says it has onboarded about 10 customers and employs roughly 125 people, while coordinating thousands of active contributors in a typical month. In the AI boom’s back office, the most expensive systems still depend on someone being awake in Hyderabad when the benchmark breaks.