Votee AI, led by CEO Pak-Sun Ting, combats the “AI data desert” by building sovereign, enterprise-grade architecture for underrepresented languages. Starting with a native Cantonese model, the company prioritises rigorous compliance and cultural preservation to empower underserved global communities.
Pak-Sun Ting, Co-founder and CEO of Votee AI, outlines how sovereign AI architecture and native language models aret reshaping enterprise governance while bringing billions of underserved language speakers into the generative AI era.
The AI revolution has a blind spot.
While large language models (LLMs) dominate headlines and enterprise adoption accelerates, most of the world’s languages remain functionally invisible to modern AI systems. According to Pak-Sun Ting, Co-founder and CEO of Votee AI, nearly 99 per cent of global languages are underrepresented in training data, leaving billions underserved. “Global LLMs are functionally illiterate outside the main languages,” he says. “Accuracy can drop below 50 per cent in some cases. We call this the AI data desert.”
It is not merely a technical shortfall. It is cultural.
“Language is the most important part of culture. If you can’t speak your own language natively in AI, you’re renting intelligence from somewhere else.”
Votee AI was built to change that.
Changing the Game
Votee’s first proof of concept was Cantonese – a language spoken by roughly 86 million people globally, yet notoriously difficult for AI systems to process.
Most existing Cantonese AI models, Pak explains, are fine-tuned overlays on Mandarin or English systems.
“Like a tourist reading from a phrasebook,” he explains, they may approximate vocabulary, but they miss tonal nuance, slang, code-switching and cultural context.
So Votee built differently.
Under the leadership of Co-founder and CTO Jacky Chan, the team developed what Pak describes as the world’s first open-source Cantonese large language model built from proprietary native corpora. They partnered with RTHK and global Cantonese communities to gather authentic data. They published the world’s first Cantonese LLM evaluation benchmark, presented at ACL 2025 in Vienna.
The result was more than technical validation. It was philosophical positioning.
“Solving Cantonese forced us to build sovereignty into the model from scratch,” Pak says. “Without native corpora, you cannot reach native fluency.”
The company even uncovered what a 16-year-old apprentice on its red team dubbed a “linguistic bypass” – harmful intent hidden within slang that global models failed to detect. A native model, trained locally, recognised the risk instantly.
“That taught us something,” Pak says. “Hong Kong’s local culture isn’t a limitation. It’s a security feature.”
From there, the road map expanded. Southeast Asia. African languages. Indigenous languages in Canada, spoken by as few as 10,000 people. Votee is preparing to launch a Global Language Data Map to visualise which languages lack sufficient AI support. “Three billion people are underserved by current AI systems,” Pak says. “If we do this right, we make that inequity visible and solvable.”
Shaping Tomorrow
Language may be Votee’s social mission, but enterprise architecture is its commercial engine.
Unlike many AI startups built for cloud-native experimentation, Votee positions itself as a sovereign, on-premise, enterprise-grade platform. Its integrated stack combines automatic speech recognition, LLMs, text-to-speech, retrieval-augmented generation (an optimisation tool) and autonomous agents in a single, governance-led infrastructure.
“In regulated industries, trust is too soft a word,” Pak says. “The standard is government enterprise grade. On-premise. Zero data leakage.”
The firm has worked with more than 200 enterprises, including government bodies and major financial institutions. It was selected for the Hong Kong Monetary Authority’s GenAI Sandbox and FSS 3.1 pilot, validating its architecture for compliant banking workloads.
Pak’s previous career explains that rigour.
Before launching Votee, he spent nearly seven years as Head of Fixed Income at Bank of China (Hong Kong), after earlier roles at Standard Chartered and J.P. Morgan. Compliance, he says, is not a checklist, it is operational DNA.
“In banking, 99 per cent is not enough. The right number is zero errors,” he says.
Votee’s Compliance AI agents monitor regulatory updates in real time, conduct automated gap analysis and generate action items, reducing regulatory research time by up to 90 per cent. Governance is not bolted on but instead is designed in.
“We say sovereignty is a feature. Trust is the outcome.”
That philosophy extends to model architecture. With over a million models in circulation globally, Pak believes the future belongs not to the best model, but to the best system for managing models.
“No single model wins every use case,” he says. “If your platform is locked into one vendor, you’re trading one dependency for another.”
Votee’s abstraction layer allows enterprises to switch models without rebuilding infrastructure, a hedge against volatility in the AI arms race.
Lessons That Last
For Pak, the pivot from institutional finance to entrepreneurship required unlearning perfection. “In banking, zero defect is the culture,” he says. “In a startup, if you wait for zero defect, you’ll never ship.”
Leaving a stable, well-compensated career was not the hardest risk. Betting on Cantonese was.
“Smart money said go after English, go after Mandarin. Big markets. Big scale. Cantonese was considered too small,” he says.
He disagreed. If the company could solve one of the most linguistically complex, under-resourced languages, it could solve any.
Today, Votee’s ambitions stretch beyond product releases. Its recent collaboration with a Hong Kong university to build an AI-driven clinical mental health simulation platform won Gold Medal with Congratulations of the Jury at the 51st International Exhibition of Inventions Geneva, awarded to only five to six per cent of global entries. This project was developed in collaboration with The Nethersole School of Nursing at The Chinese University of Hong Kong, ACTuWISE Limited, and Votee AI.
For Pak, such recognition signals a shift from feature-building to structural problem-solving. “When people come to you with pain points, not feature requests, you know you’re solving something real.”
Leadership, he says, is about preventing bottlenecks. He hires executives stronger than himself in their domains. He protects two hours a week of uninterrupted thinking time. His calendar, by design, looks underbooked.
“The strategic founder’s calendar looks empty on purpose,” he says. “The busy founder’s calendar is often a confession of weak delegation.”
And for finance professionals contemplating entrepreneurship?
“Don’t do it,” he says bluntly, unless the conviction remains even if no one is watching, funding or applauding.
This, he insists, is a long game.
Five years from now, Pak envisions enterprise AI agents operating in coordinated swarms – generating audit trails, stress-testing compliance scenarios and training clinicians in infinite simulated environments. But even as AI scales, he believes something fundamental must remain human.
“The agent is a colleague,” he says. “But accountability stays with us.”
If his career’s legacy is written in the language of code, Pak hopes it will also be written in the language of preservation. “If one day Cantonese understands us as deeply as English,” he says, “and a student in Jakarta or an Indigenous community in Canada can speak to AI in their native language, then we’ve done our job.”
In an industry obsessed with scale, Votee AI is betting on something more durable: sovereignty, architecture and the power of being understood.