From Sydney to Seoul, these AI startups are looking beyond one-size-fits-all business model
The AWS Generative AI Accelerator is supporting global startups to move beyond digital hype, creating specialised, embodied, and vocal intelligence for real-world legal, robotic, and enterprise applications.
The quest for the next leap in artificial intelligence (AI) has moved beyond the digital confines of chat boxes and into the high-stakes reality of courtrooms and factories.
In Seattle and Las Vegas, a group of forty hand-picked startups were gathered for the 2025 Amazon Web Services (AWS) Generative AI Accelerator, a programme designed to turn early-stage innovation into global infrastructure.
These companies, spanning from Brazil to South Korea, represent a shift from general-purpose AI toward highly specialised tools that solve deep-seated economic and professional problems.
A string connecting these founders is a move away from the one-size-fits-all approach of early generative AI. While the world spent the last two years marveling at tools that could write poetry or code, this cohort is focused on production-ready solutions that meet the rigid security and accuracy standards of the enterprise world.
Whether it is a robot learning to navigate a crowded convenience store or a voice assistant managing complex banking transactions, the focus has sharpened onto specific, narrow use cases where AI can outperform humans in both speed and reliability.
In conversations with YourStory, three leaders within this global cohort explain how they are redefining the competitive moat through specialised expertise, physical embodiment, and a rethink of human-machine interaction.
New moat
For Harry Raworth, Co-founder and CBO of Mary Technology, the future of the legal profession lies in extreme specialisation rather than general assistance. Based in Australia, his company focuses on a single, arduous task which is extracting objective facts from mountains of legal documentation.
In a typical litigation case, a junior lawyer might be tasked with reviewing ten thousand pages of medical records or emails to build a timeline. This process is notoriously slow and prone to human error, especially during the tenth hour of review. By using large language models to index these facts, Mary Technology allows lawyers to instantly locate specific events, such as every neck injury mentioned across hundreds of different files.
Raworth explains that the traditional competitive advantages in tech are shifting. He believes that trying to build a tool that does everything for everyone is a mistake in the current climate. Instead, the startup’s strategy is to be “deeply incredible” at one specific function, which provides a level of quality that generalist tools cannot match.
“A lot of competitors are trying to offer a little bit of everything, but in doing so, they are not truly exceptional or deeply outstanding at any one use case,” Raworth notes.
By focusing on this narrow application, the company has found rapid adoption among over a hundred law firms, ranging from small practices to global giants.
This narrow focus also addresses the historic skepticism of the legal industry toward new technology. While law firms were once anti-technology, the efficiency gains of AI have made adoption a matter of survival.
Clients are no longer willing to pay for forty hours of manual document review when a specialised tool can assist in five or six hours.
In this new landscape, Raworth identifies a different kind of protection for startups. He says that the new moat in the world of AI is velocity and speed, meaning the ability to innovate faster than the large, established players who are often slowed down by their own generalist ambitions.
Physical singularity
While Raworth organises the digital facts of the past, Junghee Ryu, Founder and CEO of RLWRLD, is focused on the physical labour of the future. Based in South Korea, Ryu is tackling what he calls the “population cliff” in East Asia, where a shrinking workforce has left factories and service industries struggling to find human staff.
The startup’s solution is a new category of technology called Physical AI, which uses vision-language-action (VLA) models. Unlike previous industrial robots that followed rigid, pre-programmed paths, VLA models allow robots to understand sensory data from the real world and react to random situations much like a human would.
This technology represents the “embodiment” of intelligence, moving AI from a screen into a physical body that can perform complex tasks. Ryu demonstrates this with a robot that uses three fingers simultaneously to open lids or pour milk, a level of dexterity that exceeds many current industry standards.
This is not just about manufacturing but also about the service sector, where companies like Lawson, a major Japanese convenience store chain, are looking to robots to restock shelves and manage items. Ryu believes we are approaching a “singularity moment” where these physical assistants will become a common sight in ordinary life.
Speaking on the timeline for this transformation, Ryu said, “Within three years, some commercial applications will begin. That means that we are now facing the singularity moment, the mass adoption of the humanoid, mass adoption of this physical AI in the market.”
He compares this coming shift to the “ChatGPT-3 moment” for robotics, where the average person will suddenly recognise the overwhelming power of physical AI to take over tough, repetitive, or “cheaper” jobs.
This will eventually lead to a split in the labour market, where humans focus on creative and complex tasks while humanoids handle the physical heavy lifting.
Human interface
The interface through which we interact with these intelligences is also undergoing a fundamental change. Akshat Mandloi, Co-founder and CTO of Smallest.ai in India, believes that the primary way humans communicate with machines will soon shift from typing to talking.
His company builds foundational voice AI models designed for high-pressure enterprise environments, such as banking and finance calling operations. These sectors have the highest urgency for automation but also the most stringent requirements for low latency, which is the delay between a human speaking and the AI responding.
Mandloi’s approach to building this voice stack is rooted in first-principles thinking, a method of breaking down complex problems into their basic truths. This philosophy extends to his hiring practices, where he ignores traditional “pedigree,” such as elite university degrees, in favour of “clarity and hunger”.
The startup’s team includes college dropouts and individuals who never attended university, yet they are training some of the highest-quality models in the industry. For Mandloi, the goal is to make human-machine interaction feel as natural and interactive as a conversation between two people.
“I feel that human machine interaction… in public settings, it will still be text based. But in personal settings, a lot of it will shift to voice led interactions,” Mandloi said.
He envisions a world where websites are voice-first, and personal assistants on our devices can instantly summarise meetings or teach us new languages through natural, interactive dialogue.
As models shrink in size and become more efficient, Mandloi believes that the real competitive advantage lies in the quality of the data and the insights a company has about a specific use case.
Infra and ethics
Despite their different sectors, the founders said that establishing enterprise trust is impossible without a “security-first” approach and robust infrastructure partners.
In the world of finance and law, security cannot be an “afterthought”. Mandloi notes that his company employs a dedicated chief information security officer to ensure compliance with the strict standards of financial institutions.
Similarly, Raworth highlights that working with law firms requires meeting intense data residency requirements, where information must be stored and processed within specific geographic borders to ensure privacy.
This is where the role of the AWS Generative AI Accelerator becomes critical. Beyond the financial support, the cloud partnership provides the technical backbone needed for these startups to scale rapidly without compromising on security.
For a growing company like Mary Technology, the ability to access the latest AI models and maintain high throughput, the volume of data processed at once, is a life-or-death matter.
Raworth explains that AWS has been a “responsive” partner, helping the startup navigate the complex technical challenges of expanding into new regions like the United States.
Mandloi also emphasises that having a partner like AWS, which is already trusted by large banks, helps “close the last mile” when deploying solutions in sensitive corporate environments.
(Cover image designed by Nihar Apte)
Edited by Affirunisa Kankudti