Deccan AI, a fast-rising competitor to Mercor in AI post-training and evaluation, has secured $25 million in an all-equity Series A to expand a high-skill expert network centered in India. The round was led by A91 Partners with participation from Susquehanna International Group and Prosus Ventures, underscoring surging demand for reliable post-training data, expert feedback, and evaluations as frontier labs push models into real-world production.
What Deccan AI Actually Does Beyond Base Model Training
Rather than training base models, Deccan AI focuses on the hard part after pretraining—teaching systems to reason, use tools, and meet enterprise-grade reliability. The company helps labs strengthen coding and agent workflows, integrate with application programming interfaces, and build reinforcement learning environments. It also runs expert-driven evaluations and delivers an enterprise product suite that includes Helix, a rigorous evaluation stack, alongside an operations automation platform designed for AI deployment at scale.
As models move beyond text to “world models” with robotics and multimodal vision, Deccan is positioning its teams to design richer tasks and more robust safety checks. Customers include Google DeepMind and Snowflake, according to the company, with roughly 10 organizations onboard and several dozen active projects at any given time.
India-Centered Talent Model for Expert AI Networks
Deccan’s core bet is that concentrating expert talent in one primary market improves quality control and speed. Most contributors are based in India, where management can standardize processes, calibrate rubrics, and iterate quickly across large projects. Around 10% of the contributor base holds advanced degrees such as master’s degrees and PhDs, though the active share can skew higher on specialized work.
Earnings vary by difficulty and turnaround, ranging from about $10 to $700 per hour, with top contributors reportedly taking home up to $7,000 monthly. That model aims to counter the pitfalls of traditional data-labeling marketplaces—where inconsistent quality and opaque incentives can derail model reliability—by emphasizing expert-led, domain-specific contributions.
India’s deep STEM bench and overlapping time zones with major tech hubs make it a natural base for this kind of work. Industry groups such as NASSCOM have repeatedly highlighted the country’s large pool of software and data professionals, and enterprises have long relied on India for complex, time-sensitive technology operations. Deccan has begun adding niche talent from other markets, including the U.S., for areas like geospatial analysis and semiconductor design.
Why Labs Are Outsourcing Post-Training Work
Post-training quality is unforgiving: small errors in reward design, edge-case prompts, or evaluation coverage can cascade into significant product failures. Frontier labs often need large volumes of high-quality feedback in days, not weeks, to iterate safely and maintain release schedules. That urgency favors specialized partners that can spin up expert teams, run red-team style evaluations, and deliver defensible metrics quickly.
The company says tolerance for mistakes in post-training is “close to zero,” which aligns with what many enterprise buyers report: measurable gains in accuracy, groundedness, and compliance often determine whether AI features ship or stall. By supplying expert feedback loops and reliable tooling, Deccan is selling speed with assurance—two things labs increasingly treat as inseparable.
Competitive Landscape and Traction in AI Post-Training
Deccan sits in a crowded field with Scale AI, Surge AI, Turing, and Mercor all chasing RLHF, data generation, and evaluation budgets. Its differentiation revolves around a “born GenAI” operating model—skipping legacy computer-vision labeling roots—and a concentrated talent footprint to keep quality consistent as complexity rises.
The company reports 10x growth over the past year and a double-digit million-dollar revenue run rate. Revenue concentration is significant—about 80% comes from its top five customers—which mirrors the broader frontier market, where a handful of buyers account for most spend. That concentration cuts both ways: it offers clear signals for product direction but heightens dependency on renewal cycles and shifting lab roadmaps.
Where the $25M Will Go: Scaling Experts and Products
Expect the capital to scale expert recruitment and deepen product capabilities. Priority areas likely include expanding Helix’s coverage across safety, tool-use fidelity, and multimodal benchmarks; automating reviewer workflows to reduce variance; and building more realistic reinforcement learning environments for agents operating across APIs and enterprise data. The company is also poised to widen its talent base for specialized domains—geospatial, chip design, biomedical coding—where credible subject-matter expertise is scarce.
Backers like A91 Partners, SIG, and Prosus have histories of supporting operationally intensive companies, suggesting a push to professionalize the expert network with stronger QA, privacy controls, and service-level guarantees that large labs increasingly demand.
What It Means for India’s Evolving Role in AI
Deccan’s strategy underscores India’s position in the AI value chain: building the expert scaffolding that turns frontier models into dependable products. While the cutting edge of model pretraining remains concentrated among a few U.S. and Chinese players, the labor-intensive, high-skill layers of evaluation and reinforcement learning are becoming India’s proving ground. If Deccan sustains quality at scale, it will not only pressure rivals like Mercor and Turing but also set a template for India-led expert platforms in safety-critical AI workflows.
The bigger takeaway: as AI moves from demos to deployment, the bottleneck shifts from compute alone to disciplined human expertise. Deccan’s $25 million raise is a wager that the next leap in reliability will be won as much in expertly managed post-training pipelines as in the training run itself.