Deccan AI, a startup that supplies expert human data for training and testing AI models, has raised $25 million in an all equity Series A led by A91 Partners, with Susquehanna International Group and Prosus Ventures joining the round. TechCrunch first reported the deal on March 25, 2026, describing the company as a Mercor competitor that recruits most of its contributors in India. The pitch is simple: the labs building frontier models need near flawless training data, and Deccan argues that one tightly managed talent pool delivers better quality than a workforce scattered across dozens of countries.
Below: the deal terms, what Deccan sells, how its India based network works, how it compares with its rivals, and what it has shipped through late September 2026. Every figure comes from the original reporting or from Deccan’s own announcements.
The round: who invested and what was disclosed
According to TechCrunch, the $25 million was Deccan’s first major funding round and was structured entirely as equity. A91 Partners, an India focused growth investor, led it. Susquehanna International Group and Prosus Ventures participated. Prosus was not new to the company: Deccan announced an undisclosed investment from Prosus Ventures in May 2025, when it described a freelancer network of more than 500,000 people.
Deccan’s own blog announcement, posted April 14, 2026, framed the raise around what it calls “superaccurate AI” and pointed the money at reinforcement learning environments, evaluation systems and enterprise AI work. No valuation was disclosed, and Deccan did not share exact revenue. It did say the business grew roughly tenfold over the prior year and runs at an annualized revenue figure in the double digit millions of dollars.
The original TechCrunch report is the best single source on the deal. It also carries the most revealing detail about risk: about 80% of Deccan’s revenue came from its five largest customers at the time of the round.
What Deccan AI actually does
Deccan works on the part of model development that happens after pretraining. Once a lab has a base model, it needs carefully written examples, expert feedback, preference rankings and hard test cases to make the model more accurate and more useful. That work goes by several names (supervised fine tuning data, RLHF, evals), and it depends on people who genuinely know the subject, whether that is tax law, organic chemistry or a tricky piece of Python.
The company groups its products into three lines:
- STARK RL environments: simulated workplaces where AI agents practice using tools, browsers, operating systems and code, with automated checks on whether they did the job correctly.
- Helix: an evaluation and monitoring suite that mixes automated scoring with human review to catch drift and failures in deployed models.
- EnterpriseOS: custom AI agents for back office and middle office operations, aimed at large enterprises and global capability centers.
At the time of the round, Deccan said it had onboarded about 10 customers and typically ran a couple of dozen active projects. Google DeepMind and Snowflake are the customers named in the TechCrunch report. Reddy told TechCrunch that tolerance for errors in post training work is close to zero, and summed up the market opportunity by saying that “quality remains an unsolved problem.”
Why India, and how the expert network works
This is the part of the story the headline is about. Deccan is headquartered in the San Francisco Bay Area (its registered office is in Mountain View, California), but its main operations hub is in Hyderabad and it also lists an office in Bangalore. Most of its contributors are in India.
Reddy’s argument, as reported by TechCrunch, is that many competitors recruit experts across more than 100 countries, while running operations in one country makes it far easier to hold everyone to the same quality bar. Deccan has started recruiting in the US as well, but only for specialist areas where India’s pool is thinner, such as geospatial data and semiconductor design.
The numbers the company shared give a sense of scale:
- More than 1 million registered contributors, including students, domain experts and PhDs.
- Between 5,000 and 10,000 contributors active in a typical month.
- Pay ranging from $10 to $700 an hour depending on the task and expertise, with top earners making up to $7,000 a month.
In December 2025, before the round, Deccan launched a branded recruiting arm called Deccan AI Experts, aimed at what it described as India’s top professionals, and said it had a bench of more than 12,000 vetted specialists ready for complex projects. Platforms in this space screen applicants before assigning work (micro1, for example, vets experts with an AI interviewer), so coders should be ready to talk through their reasoning; our guide on how to explain your coding solution in an interview covers the same skill reviewers look for.
How Deccan compares with Mercor, Surge AI, Scale AI, Turing and micro1
Deccan is a small player in a market that has attracted enormous capital. TechCrunch placed it alongside Scale AI, Surge AI, Turing and Mercor, all competing to sell labeling, evaluation and reinforcement learning services to the same handful of labs. micro1 belongs on the list too. Here is how the field looked in September 2026, using only figures from original reporting or company announcements.
| Company | What it sells | How it finds experts | Latest verified funding news |
|---|---|---|---|
| Deccan AI | Post training data, evals (Helix), RL environments (STARK), enterprise agents | Network concentrated in India, plus US recruiting for niche fields | $25M Series A led by A91 Partners, March 2026; valuation not disclosed |
| Mercor | Expert marketplace that supplies specialized training data to AI labs | Recruits PhDs, lawyers, bankers, scientists and programmers | $350M Series C at a $10B valuation, October 2025; reported in July 2026 to be in talks at $20B |
| Surge AI | Human data for RLHF and model evaluation | Large contractor workforce; founded in 2020 by Edwin Chen | Bootstrapped; reported in 2025 to be discussing its first outside investment |
| Scale AI | Data labeling and evaluation at very large scale | Large global contractor workforce | Meta invested about $14.3B for a 49% stake, June 2025 |
| Turing | Coding, reasoning and multimodal post training data for frontier labs | Network of vetted software engineers dating to its 2018 founding | $111M Series E at a $2.2B valuation, 2025 |
| micro1 | Expert data, RL environments, robotics data | Screens experts with an AI interviewer | More than $100M at a $4B valuation, reported September 22, 2026 |
The gap in scale is obvious. Mercor, per the TechCrunch report on its funding talks, was running at about $2 billion in annualized revenue by mid 2026, a figure that dwarfs Deccan’s double digit millions. The same report noted that Mercor suffered a data breach in April 2026 and that several contract workers filed lawsuits, a reminder that fast growth in this sector comes with real operational risk.
Deccan’s differentiation is therefore about positioning rather than size. Reddy described Deccan to TechCrunch as a “born GenAI” company, meaning it focused on higher skill generative AI work from the start rather than adapting an older labeling business. Its single country model is the other bet. Whether that is an advantage or a ceiling depends on how many languages, jurisdictions and niche specialties the big labs end up needing.
What Deccan has done since the round
We found no reports of a new funding round, acquisition or leadership change at Deccan AI between the March announcement and September 23, 2026. The company’s own blog does show steady product work:
- April 14, 2026: Deccan published its Series A announcement, restating the investor lineup and its focus on accuracy.
- May 25, 2026: the company introduced STARK RL environments (the name stands for Simulated Environments, Trajectories and Agentic RL). The environments cover tool use, browser use, computer use and coding, and each task is checked by five independent verifiers, including a database check that confirms an agent’s changes actually happened.
- September 11, 2026: Deccan released CaptionBench, a human reviewed benchmark for video captioning. Reviewers assessed 1,144 captions from six frontier models across 196 short clips. The headline finding was that models with nearly identical leaderboard scores failed in very different ways, and that all six struggled with describing camera movement.
A benchmark like CaptionBench doubles as a sales pitch: it shows labs the kind of expert review Deccan can run, which is exactly the paid work of finding where a model fails and producing data to fix it.
What it means for labs, enterprises and experts
For AI labs, Deccan is one more supplier in a market where buyers spread work across vendors, and its STARK and Helix products target the agent and evaluation work that has become central to post training.
For enterprises, EnterpriseOS is the more interesting signal. Several expert data companies are trying to turn what they learned serving labs into products for ordinary businesses, and Deccan explicitly named finance, compliance, support and internal operations teams as targets.
For experts, the practical takeaway is that this is project work with uneven demand. The gap between 1 million registered contributors and 5,000 to 10,000 active ones in a given month tells you most sign ups will not get steady hours. Freelancers who already juggle clients should treat it as one income stream among several.
Frequently asked questions
How much did Deccan AI raise, and who led the round?
Deccan AI raised $25 million in an all equity Series A, its first major funding round. A91 Partners led it, and Susquehanna International Group and Prosus Ventures participated. TechCrunch reported the deal on March 25, 2026. Prosus had already backed the company with an undisclosed investment announced in May 2025. No valuation was made public.
Who founded Deccan AI?
Rukesh Reddy is the founder and CEO. His company profile lists more than 15 years at Citi, Monitor Group and J.P. Morgan, plus degrees from IIT Bombay and IIM Ahmedabad. TechCrunch reported that Deccan was founded in October 2024 and employed about 125 people when the Series A was announced in March 2026.
Why is Deccan AI called a Mercor competitor?
Both companies recruit skilled professionals to create training and evaluation data for AI labs. The difference is scale and sourcing. Mercor is far larger and recruits professionals across many fields, while Deccan concentrates most of its contributor network in India, arguing that one country makes quality control easier. Both sell to the same frontier labs and enterprises.
How much do Deccan AI contributors earn?
According to TechCrunch, contributors earn between $10 and $700 an hour depending on the task and the expertise required, and top earners make up to $7,000 a month. Only 5,000 to 10,000 of more than 1 million registered contributors are active in a typical month, so steady hours are not guaranteed.
Which companies use Deccan AI?
TechCrunch named Google DeepMind and Snowflake as customers, and Deccan’s website also displays Adobe among its clients. At the time of the round the company said it had onboarded about 10 customers and was running a couple of dozen active projects, with most revenue coming from its five largest accounts.
The bottom line
Deccan AI’s $25 million Series A is modest next to rivals valued in the billions, such as Scale AI, Mercor and micro1, but it funds a clear thesis: expert data for post training is a quality problem first, and a concentrated India based workforce is the way to solve it. The company has followed the raise with real product releases, from STARK RL environments in May to CaptionBench in September.
The open questions come from its own numbers: heavy reliance on five customers, a wide gap between registered and active contributors, and rivals with far deeper pockets. For now, Deccan is the name to watch among smaller expert data vendors.
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.

