Dominic Jainy stands at the intersection of emerging technology and industrial application, offering a seasoned perspective on the rapid evolution of artificial intelligence within the Asia-Pacific region. As an IT professional specializing in machine learning and foundational models, he has closely monitored Singapore’s Sapiens Technology as it transitioned from a promising local startup to a major regional powerhouse. This conversation highlights the strategic moves that led to their inclusion in the 2026 Forbes Asia 100 to Watch list and the technical precision required to compete on the global stage with the Agnes AI ecosystem. We explore the balance between maintaining high-performance benchmarks and ensuring that frontier-grade AI remains economically viable for the developers who drive innovation.
The conversation touches upon the significance of in-house model training for regional competitiveness and the role of independent benchmarking in establishing developer trust. We delve into the economic strategy behind pricing video generation for production and the logistical scale of processing trillions of weekly tokens. Finally, we examine the technical hurdles of multi-modality and what the future holds for the Asia-Pacific AI landscape.
Sapiens Technology was recently highlighted as one of only 15 Singaporean companies in the Enterprise Technology category on a major regional watchlist. What specific innovations in your in-house training processes helped you stand out, and how has this recognition impacted your growth strategy?
The decision to train our foundation models entirely in-house across text, image, and video was a fundamental shift that allowed us to move beyond the limitations of third-party wrappers. Being named one of only 15 Singaporean companies in the Enterprise Technology category on the Forbes Asia 100 to Watch validates our focus on vertical integration and localized innovation. This recognition has sharpened our growth strategy, pushing us to expand our infrastructure to better serve the 16 countries represented on this year’s list. We are now doubling down on providing a stable, high-performance base that ensures Singapore remains the second most-represented market in the region for tech innovation.
With Agnes 2.5 Pro Alpha ranking #9 on the Intelligence Index and scoring 58.8 for coding, how do these independent benchmarks influence developer trust? Could you walk us through the technical hurdles of maintaining high performance across text, image, and video modalities simultaneously?
When Agnes 2.5 Pro Alpha debuted at #9 out of 153 models on the Intelligence Index, it provided the developer community with a necessary “truth anchor” that goes beyond marketing claims. Scoring a 58.8 on the Coding Index is particularly significant because it demonstrates that our model isn’t just a generalist, but a tool capable of high-level logic and syntax. The technical hurdle of maintaining this performance across text, image, and video simultaneously is immense because each modality competes for the same underlying compute architecture during training. We have to carefully manage data weights to ensure that the rich visual data required for video doesn’t degrade the precision and linguistic nuances needed for high-level coding tasks.
Providing frontier-grade AI at price points like $0.30 per minute for video generation suggests a focus on production-ready scalability. How do you balance the high costs of infrastructure with the goal of price parity, and what metrics indicate that this approach is sustainable for your ecosystem?
Offering frontier-grade AI at a price point of $0.30 per minute for video generation is a calculated move to achieve what we call AI parity. We balance these high infrastructure costs by optimizing our model’s inference efficiency, which allows us to provide pricing that is actually ready for commercial production rather than just short demos. The sustainability of this approach is measured by our high ranking on the Artificial Analysis leaderboards combined with the growing number of developers moving away from more expensive, less transparent alternatives. We believe that if builders cannot afford to use the tools, the benchmarks themselves become irrelevant to the industry’s actual progress, so our primary metric for success is the ratio of cost to benchmark performance.
Your ecosystem currently processes over 8 trillion tokens per week while offering free API access for testing. What specific steps must a developer take to move from an initial trial to full-scale production, and how do you manage the resource demands of such high volume?
Currently, our ecosystem is processing over 8 trillion tokens per week, which is a testament to the massive adoption we are seeing across different sectors. For a developer to move from our free API trial to full-scale production, they simply need to transition through our platform’s integration portal, which is designed to handle rapid scaling without technical friction. We manage these massive resource demands by utilizing a distributed infrastructure that scales dynamically based on real-time load across the Agnes ecosystem. This high volume of traffic actually helps us refine our models further, as the diverse range of queries provides a rich feedback loop that we use for continuous optimization of our foundation layers.
Building foundation models across multiple modalities is a resource-intensive endeavor. Can you share an anecdote about a specific challenge your team faced while developing Agnes-Video-2.5, and how did you ensure the final output met the rigorous standards of third-party evaluation?
Developing Agnes-Video-2.5 was one of the most resource-intensive projects we have ever undertaken, especially when we hit the hurdle of ensuring temporal consistency in generated scenes. One specific challenge involved maintaining the structural integrity of objects over several seconds of footage, a common failure point that we had to solve through repeated iterations of our training datasets. To ensure we met the rigorous standards of third-party evaluation, we ran thousands of internal tests before ever submitting our models to the Text-to-Video Leaderboard. Our founder, Bruce Yang, often emphasizes that a rank only matters if it solves real-world problems, so we focused heavily on the practical output quality that developers expect for commercial video.
What is your forecast for the future of multi-modality AI development in Singapore and the broader Asia-Pacific region?
My forecast for the future of multi-modality AI in Singapore is one defined by hyper-specialization and regional leadership as companies seek more efficient ways to integrate text, image, and video into their workflows. We are moving into an era where the focus will shift from simply generating content to creating deeply integrated enterprise solutions that operate at scale. Singapore is uniquely positioned to lead the Asia-Pacific region because of its robust regulatory environment and its status as a primary hub for tech talent. I expect that within the next few years, multi-modal AI will become the invisible backbone of the entire regional enterprise sector, with foundation models like Agnes setting the standard for both performance and affordability.