The invisible layer powering frontier AI
When you talk to a cutting-edge AI model and it responds with nuance, cultural awareness, and something approaching common sense - that didn't happen because the model figured it out alone. Somewhere in the pipeline, thousands of human experts read, labeled, rated, and refined what that model knew. Eric Zhang runs the company that builds that pipeline.
As CEO of Thoth AI, Zhang leads a global AI data solutions company headquartered in Singapore with R&D operations now rooted in San Francisco's Silicon Valley. Thoth AI operates in 170+ countries, works across 200+ mainstream languages, and serves some of the world's most demanding AI labs - helping them build production-grade models from raw data into reliable, deployable systems.
The company's name is not an accident. Thoth was the ancient Egyptian god of knowledge, wisdom, and writing - the original record-keeper of what intelligence means. Zhang's company occupies the same role for modern AI: teaching machines what it means to know something, say something accurately, and treat people fairly across every language and culture on earth.
"Empowering innovators everywhere to build AI that is powerful, responsible, and globally adaptable."
Zhang runs the company from California while overseeing a headquarters perched on Level 42 of 6 Battery Road in Singapore - one of the island city-state's most prestigious business addresses. That geographic duality is the model itself: Silicon Valley's innovation intensity, Southeast Asia's linguistic and cultural depth, fused into something neither could build alone.
Thoth AI's team of nearly 880 professionals spans offices across the Philippines, Indonesia, Malaysia, Vietnam, Spain, and Portugal - each chosen not just for cost efficiency but for genuine domain expertise. Healthcare annotation teams with clinical context. Gaming AI teams who actually play. Finance data teams who understand the difference between a hedge and a hedge fund.
This is the part of the AI industry the press rarely covers: the annotators, the quality reviewers, the human-in-the-loop specialists who make the difference between a model that sounds plausible and one that's actually correct. Zhang has made that workforce the core product, not a back-office cost.
Six disciplines, one invisible backbone
Thoth AI's service architecture is a blueprint for what it takes to get a language model from experimental lab prototype to production-grade system that real people trust with real decisions.
Data Collection & Annotation
Images (2D & 3D), video, 3D point cloud, text, audio - labeled with precision by domain experts. The raw material that every AI model runs on.
Generative AI & RLHF
Reinforcement Learning from Human Feedback using expert-in-the-loop methods. Where models learn the difference between "technically correct" and "actually useful."
Model Evaluation
Rigorous evaluation frameworks for LLMs, VLMs, and multimodal systems. Catching hallucinations, bias, and edge cases before they reach users.
Trust & Safety
Content moderation, fraud prevention, and risk management at scale. The systems that keep AI deployments from causing harm in the real world.
Multilingual Customer Service
Native-speaker teams across 60+ languages. Because "multilingual support" that isn't actually fluent isn't support at all.
Global System Implementation
Deployment support, product testing, and operational infrastructure for AI systems going live in complex real-world environments.
Why Zhang moved R&D to Silicon Valley
In 2025, Zhang announced that Thoth AI was opening a dedicated R&D hub in the San Francisco Bay Area. From the outside it might look like a typical tech company following the orbit of capital and talent. From the inside, it's a deliberate bet.
The frontier AI labs that Thoth AI serves - the ones building foundation models that will define the next decade - are concentrated in San Francisco. Being in the same timezone, the same conference rooms, and the same hallways as those teams is not a minor logistical advantage. It changes the nature of the partnership.
Zhang described the move as "merging worldwide capabilities with Silicon Valley's innovation ecosystem." What he was really doing was completing the circuit: Southeast Asia's deep bench of skilled annotators and domain experts at one end, Silicon Valley's model builders at the other, with Thoth AI as the conductor in the middle.
"Next-generation data solutions, generative AI, and responsible innovation - built at the intersection of global scale and Silicon Valley precision."
The R&D hub focuses on three areas that are defining the competitive landscape in AI right now: advanced data collection methods for next-generation models, generative AI techniques that improve model quality, and responsible innovation frameworks that allow AI to be deployed safely across markets with different cultural norms and legal requirements.
Gaming is AI's most demanding live laboratory
At Gamescom Asia x Thailand Game Show 2025 in Bangkok, Zhang made an argument that the gaming industry doesn't often hear from AI infrastructure companies: that game environments are not just a use case for AI - they're the most rigorous test bed AI has.
"Millions of people interact with AI every single day" in games, Zhang has said. Those interactions aren't scripted, controlled, or repeatable. A player can ask an NPC anything, approach a situation from an unexpected angle, or test the emotional responsiveness of a virtual companion in ways no lab benchmark anticipated. If the AI breaks, the player notices immediately.
Thoth AI's positioning in gaming is a case study in how the company thinks about every vertical: the game industry needs smarter NPCs, seamless multilingual worlds, content moderation at massive scale, and AI companions that don't just respond - they relate. Each of those requirements maps directly to Thoth AI's core capabilities: RLHF for behavior refinement, multilingual annotation for localization, trust and safety pipelines for content, and human feedback loops for emotional intelligence training.
"AI should make our world more alive, not less human."
- Eric Zhang, CEO of Thoth AI"Gaming is the front line where millions interact with AI every day."
- Eric Zhang at Gamescom Asia 2025"Empowering innovators everywhere to build AI that is powerful, responsible, and globally adaptable."
- Thoth AI Mission"AI that can speak, think and dream across languages and cultures."
- Eric ZhangBuilding at global scale
What "responsible AI" actually looks like at scale
Most tech executives talk about responsible AI. Fewer have built the operational infrastructure to deliver it across 170 countries in 200 languages. Zhang's bet is that responsible AI isn't a values statement - it's an engineering problem that requires human expertise at every layer.
Thoth AI's approach to AI safety is structural, not rhetorical. Their trust and safety services include content moderation, bias detection and correction, AI fairness assessment, and model reliability validation. Their RLHF workflows specifically address the problem of AI models that perform well on benchmarks but fail in the real world - the gap between "the model answered correctly" and "the model behaved correctly in context."
The multilingual angle is not marketing decoration. When you're training an AI model that will be deployed in Indonesia, Vietnam, or Portugal, you need annotation done by people who are culturally native to those contexts - not just fluent in the language, but fluent in the assumptions, sensitivities, and idioms that the text carries. Thoth AI is one of the few companies that has built that workforce at production scale.
Zhang's aspiration, stated plainly: to make Thoth AI the world's most trusted AI data infrastructure company. The kind of partner that AI teams call when they need to move from "this works in the lab" to "this works in the world."