Deccan AI

A room of people, very politely, telling a machine it is wrong.

On any given Tuesday in Mountain View, somewhere inside the offices at 257 Castro Street, a domain expert is reading a model output and writing a precise note about why the third bullet point would get a junior engineer fired. They are not, technically, training the model. They are doing something quieter and more expensive: producing the human-verified row of data that, multiplied a few hundred thousand times, will eventually teach the model not to do that again.

This is Deccan AI's whole business. It is unglamorous, it is mechanical, it is - increasingly - load-bearing. The world's frontier labs spent the last three years racing each other on parameter counts. Now they have all noticed the same thing: the bottleneck moved. The new bottleneck is people who can read a model answer about a Postgres query, or a LaTeX proof, or a robotic arm motion plan, and say with confidence what "better" actually means.

Frontier models ran out of internet.

By late 2024 the dirty secret of the model labs was an open one. The crawled web had been crawled. The benchmarks had been gamed. Every additional point on MMLU was costing more, and the marginal token of pre-training data was, statistically, an SEO blog about banana bread. The next round of capability gains - reasoning, agents, code that compiles in production - was not going to come from scraping harder.

It was going to come from post-training. From SFT. From RLHF. From environments where an agent could try, fail, and be corrected by someone who actually knew what failure looked like. Which meant the labs needed, urgently and at scale, the one input they had spent the previous decade pretending they didn't need: humans, expensive ones, with opinions.

The annotation industry that existed at the time was built for an earlier era. Bounding boxes on cats. Sentiment tags on tweets. It was not built for "evaluate whether this agent correctly used the API to refund the customer," and it certainly wasn't built for "explain in writing why this generated proof has a subtle gap on line four." There was a gap. Deccan AI walked into it.

A finance guy decides AI's biggest problem is a recruiting problem.

Rukesh Reddy is, on paper, a strange person to be running a company that supplies data to GPT-class labs. Fifteen years across Citi, Monitor and JPMorgan. IIT Bombay undergrad, IIM Ahmedabad MBA. Most recently, Head of Growth at an Indian wealth manager. He is not, in other words, a typical founder of an LLM-adjacent startup.

What he was, very specifically, was an operator who had spent his career watching how large enterprises spec, source and quality-check expert labor. When he looked at the AI data market in 2023, he saw something almost embarrassingly familiar: a procurement problem dressed up in a hoodie. The labs needed PhDs, senior engineers, working lawyers, math olympiad medalists - and they needed them organized into workflows with audit trails and quality controls. That is a business. It just was not one that anyone was running well.

So he built one. Around him he pulled an org chart that reads less like a startup and more like a consulting bench: Jagdeep Sahni from AWS and Intuit running revenue, Purva Bhandari from Google and McKinsey on strategy, Samir Janveja from Google on strategic accounts, Mahesh Reddy formerly head of editors at HackerRank running engineering, Sajo Mathews ex-Samsung director of engineering running ML. The vibe is operator-heavy on purpose. Frontier labs do not buy from cowboys.

A platform that linters can't lie to.

What Deccan AI sells is, depending on which slide you are looking at, a service, a platform, or a network. In practice it is all three, glued together in a way that is hard to copy.

The clever bit is the platform layer. Most annotation tools are essentially form builders with a queue attached. Deccan's tool ships with a built-in code compiler, a LaTeX editor and a small army of LLM-based validators that pre-flag bad submissions before a human ever sees them. The result is a workflow where the cheap parts of quality control are automated and the expensive parts - the actual human judgment - get spent on the questions that need it.

Three years, one straight line.

Numbers that don't need a footnote.

Three years in, the company books revenue from a customer list it is mostly not allowed to discuss in writing - except to say that it includes a majority of the so-called "Magnificent 7." The public reference is Snowflake, whose principal engineer for RAG and agentic evaluation was willing to put a quote on the record.

Pristine data, or no data at all.

The company's stated mission is to "empower companies with the highest-quality, human-verified datasets, becoming the trusted partner for superior AI model training." The vision goes a step further: to set the global standard for quality, integrity and innovation in AI data. Read in the wrong tone these come off as boilerplate. Read in the right one - against the backdrop of an industry where model outputs are routinely confidently wrong - they sound almost defensive.

There is a kind of moral logic to the company that is easy to miss. If the next decade of AI is going to involve agents that take real actions in the real world - moving money, filing legal documents, dispatching code into production - then the data those agents are trained on has to be auditable. You should be able to point at the example that taught the model how to refund a customer and say: a person wrote this, a second person checked it, and here is their thinking. Deccan AI sells exactly that paper trail.

The boring layer is about to get extremely interesting.

The Series A money will go where the company has been going anyway: deeper into agentic workflows, more RL environments, more coverage of the long tail of specialty domains - finance, legal, KYC, healthcare, the physical world. The strategic question for Deccan AI is whether to remain a high-trust services-and-platform hybrid, or whether to lean harder into the software side and try to become the system of record for post-training data the way Snowflake became the system of record for analytics.

Either path runs into the same competitive set: Scale, Surge, Mercor, Invisible. None of them are sleeping. What Deccan has, that the others don't, is a quieter brand and an operator bench that does not embarrass itself in front of enterprise procurement. That sounds modest. In this market it is the moat.

A note, written by a person, that the model will eventually obey.

Back in Mountain View it is now a Tuesday afternoon and the senior annotator has finished her note about the third bullet point. It is two sentences long. She tags it with a domain code, attaches a reference link to the relevant API documentation, and submits it into a workflow that will route it to a second reviewer, then a third, then a project manager whose only job is to make sure that this exact correction is captured in a way the customer's training pipeline can ingest.

Three weeks from now, the next version of a model you have heard of will ship. Somewhere inside it, that two-sentence note will have done its small, anonymous, structural work. The model will be a little less wrong. Nobody outside the room will know why. That is the point. That is the whole company.