Deccan AI Raises $25 Million to Build More Reliable Post-Training Data

MOUNTAIN VIEW, Calif. - Deccan AI has raised $25 million in Series A financing to expand its post-training data, evaluation and enterprise AI business. A91 Partners led the all-equity round, with Susquehanna International Group and existing investor Prosus Ventures participating. The capital injection comes as the artificial intelligence industry shifts its focus away from the raw computational power required for initial model training and toward the meticulous refinement necessary to make these systems reliable for commercial deployment. Deccan AI operates at this intersection, providing the human-in-the-loop expertise and software infrastructure required to calibrate large-scale models after they have ingested their primary datasets.

The central thesis of Deccan AI’s operation is that foundation models are not finished when their initial training run ends. While a model’s primary pre-training phase allows it to learn the general structure of language or imagery, it often lacks the specific guardrails, factual accuracy, and domain-specific nuance required for high-stakes enterprise applications. Developers still need specialized data, human feedback, evaluations and reinforcement-learning environments to improve accuracy and make models useful in specific industries. This post-training phase has become the primary battleground for model labs seeking to reduce hallucinations and improve the safety of their outputs.

Deccan supplies that layer through a network of experts and software products built for model labs and enterprise customers. By integrating human intelligence with automated evaluation workflows, the startup aims to provide a more rigorous testing ground than traditional crowd-sourced labeling services. This specialized approach is increasingly necessary as models move into complex fields like medicine, engineering, and enterprise software development, where a generalist’s feedback is no longer sufficient to identify subtle errors in a model's logic or technical output.

The company has a Silicon Valley headquarters and a large operations presence in Hyderabad, India. That structure gives Deccan access to technical and domain specialists while keeping it close to major U.S. AI customers. This dual-presence strategy is typical of modern infrastructure firms that require both the high-level engineering talent found in Northern California and a scalable workforce of subject matter experts in major global technology hubs. The Hyderabad office serves as a nexus for the company’s vast network of evaluators, many of whom hold advanced degrees or professional certifications in specialized technical fields.

Its business also reflects a broader shift from low-cost data labeling toward higher-skill work involving coding, science, law, finance and model evaluation. In the early days of machine learning, data preparation largely involved simple tasks like drawing bounding boxes around cars in photographs or identifying objects in videos. However, the rise of generative AI has fundamentally altered the labor requirements of the industry. Modern model evaluation requires contributors who can write complex software code, solve graduate-level physics equations, or interpret the nuances of contract law to determine if a model is responding appropriately to a prompt.

Deccan said it will use the financing to invest in research and development, strengthen infrastructure for enterprise deployments and expand its work in robotics and other data-intensive fields. The expansion into robotics is particularly significant, as autonomous physical systems require massive amounts of precisely labeled sensor data and human intervention to learn safe behaviors in unpredictable environments. By building software specialized for these data-intensive domains, Deccan hopes to entrench itself as a critical vendor for companies moving beyond simple text-based chat interfaces and into more complex physical or multimodal AI applications.

The company competes with a growing group of platforms that source expert talent for AI training and evaluation. As the demand for high-quality data has surged, several well-funded players have entered the market, each vying for the same limited pool of PhDs, lawyers, and developers who can provide the necessary feedback loops. Analysts have noted that the market for 'ground truth' data is becoming increasingly competitive, with startups needing to differentiate themselves through superior software tools, better quality control, and faster turnaround times for model labs that are operating on aggressive development cycles.

The financing comes at a time when the sustainability of AI scaling is under heavy scrutiny. Many industry observers argue that the industry is approaching a 'data wall,' where the supply of high-quality public text on the internet has been exhausted. This makes the creation of proprietary, high-quality post-training data even more valuable. Deccan’s position suggests that the next frontier of AI development will not be determined simply by who has the most GPUs, but by who has the most sophisticated pipeline for distilling human knowledge into a format that models can ingest to improve their reasoning capabilities.

For enterprise customers, the barrier to AI adoption is often a lack of trust in model reliability. Deccan AI’s software infrastructure aims to solve this by providing a framework for continuous evaluation. As enterprises deploy AI agents into production, they require constant monitoring to ensure that performance does not drift and that the models do not begin to output incorrect or biased information. By providing both the expert labor to set the initial benchmarks and the software to track performance at scale, Deccan is attempting to offer an end-to-end solution for corporate model governance.

The main risk is that model developers increasingly automate parts of the data pipeline or bring sensitive work in-house. Major labs like OpenAI and Google have expressed interest in 'synthetic data,' where one AI model is used to train or evaluate another. If the industry moves toward a purely automated model for training and refinement, the demand for human-expert networks could diminish. Furthermore, as AI models become more foundational to a company’s core intellectual property, some may choose to build their own internal evaluation teams rather than outsourcing to third-party platforms like Deccan.

Deccan's opportunity depends on proving that its experts and evaluation systems deliver higher-quality outcomes than generic labor marketplaces or synthetic data alone. The company argues that while synthetic data can help with scale, it often lacks the edge-case knowledge and creative problem-solving capabilities of a human expert. For mission-critical tasks where the cost of an error is high, human-verified data remains the gold standard. Deccan must continue to innovate on its software layer to ensure that it is not merely a staffing agency, but a technology company that provides measurable improvements in model performance.

The Series A gives it capital to make that case at a larger scale. With $25 million in new funding, the company can expand its operational footprint and invest in the specialized tooling required to support increasingly complex model architectures. The involvement of A91 Partners and Susquehanna International Group provides the firm with a stable base of institutional support as it navigates a rapidly shifting landscape. As the first wave of generative AI hype transitions into a more sober phase of enterprise implementation, the focus on data quality and evaluation is likely to remain a central theme for the broader technology sector.