Opinions expressed by Digital Journal contributors are their own.
Ask most people what responsible AI looks like and you will get a list of principles. Ask Harsh Singhal and you will get a system architecture. That difference, between stated principles and operational systems capable of enforcing them, has defined the central problem of his career. Across roles at LinkedIn, Netflix, Adobe, Koo, and now Glean, Singhal has worked in environments where responsible AI was an engineering requirement backed by real consequences, and the systems had to hold up under the full pressure of production.
His most visible test came at Koo, where he was handed one of the hardest applied AI problems in the Indian technology sector and asked to solve it at social media scale.
The scale of the problem
Koo was designed to serve Indian users in their own languages, and that vision required a content safety infrastructure capable of handling hate speech, harassment, and harmful content across Hindi, Tamil, Telugu, Kannada, Bengali, Marathi, and several other languages, all simultaneously, all in real time, in a user environment where people mixed languages within single posts and wrote in transliterated forms that no standard classifier had been trained to handle. When Singhal joined as Senior Director and Head of Machine Learning in 2021, the team numbered three engineers and the platform was growing rapidly, having gained millions of users during a period of heightened tension between the Indian government and Twitter. The moderation infrastructure needed to scale as fast as the user base, which meant building new systems while keeping existing ones running across a linguistic environment more complex than almost any social platform had previously attempted to govern.
“Building a multilingual safety system is not a translation problem,” Singhal said. “Translating your content policies into other languages is the straightforward part. The hard part is teaching a system to understand how people actually express hostility, manipulation, or harm in those languages, including the ways that differ completely from how the same intent appears in English.”
Building the team and the technology
Singhal scaled the machine learning team from three to twenty engineers, creating a multidisciplinary group spanning data science, machine learning engineering, and MLOps, a deliberate organizational build designed to match the technical complexity of the problem. On the technical side, his most significant contribution was leading the development of KooBERT, an open-source multilingual transformer model purpose-built for Indian-language content across more than 20 languages. Existing multilingual models had been trained primarily on formal written text and performed poorly on the code-mixed, transliterated, script-variable content that characterized Koo’s user base, and KooBERT was engineered specifically for those conditions.
Making it open-source extended its potential reach beyond Koo, giving researchers and engineers working on Indian-language NLP a purpose-built resource with real production provenance. Alongside KooBERT, Singhal led the adoption of Meta’s LLaMA models, fine-tuned for multilingual toxicity detection, positioning Koo as one of the first social platforms globally to deploy fine-tuned large language models for real-time content safety applications. Mayank Bidawatka, Co-founder of Koo, noted that this work “left a lasting legacy in India’s digital transformation, demonstrating how responsible AI can empower local communities while setting new standards for scalable, ethical technology in social networking.”
The content moderation infrastructure his team built supported detection and removal of harmful content across the platform’s full linguistic range, while the recommendation systems they built simultaneously, including Feed Ranking, Semantic Search, Multilingual Topics, and People You May Know, served personalized content to users across all supported languages. “You cannot separate the safety work from the personalization work,” Singhal said. “The same underlying language understanding that powers a good recommendation also powers a good moderation decision. We built the foundation to serve both, because the platform needed both to work at the same time.”
What 60 million users teaches you about responsible AI
Koo reached over 60 million users at its peak, a scale at which the consequences of safety system failures become visible and public very quickly. Managing content moderation across that user base, in ten languages, with the real-time latency requirements of a social feed, produced practical lessons about responsible AI that no research environment replicates. Academic benchmarks for hate speech detection are typically constructed from labeled datasets that reflect the distribution of content at a particular point in time, in a particular language register, and production content drifts, evolves, and responds to current events in ways that benchmark performance cannot predict. Singhal’s approach at Koo involved continuous model updating, active monitoring of system outputs, and feedback mechanisms that kept the moderation systems aligned with actual platform content rather than historical training data.
Another lesson from operating at that scale concerns explainability. A moderation system that produces correct outputs but cannot explain its reasoning creates operational and user trust problems, particularly in environments where appeals processes and human review must function alongside automated decisions. Singhal’s team worked on systems that could support that review layer, providing enough transparency for moderators and users to understand why a decision had been made, a design principle that carries directly into his current enterprise work.
Responsible AI as an engineering discipline
At Glean, Singhal applies the same principles to enterprise AI governance. His work on sensitive content detection uses contextual signals, permissions data, and enterprise graph information to identify genuinely sensitive content in unstructured enterprise data, combining technical controls with policy logic so that governance operates at system speed.
His more recent work has focused on technical systems for enterprise-aware data security, AI agent assurance, and adaptive enterprise intelligence, all aimed at making governance workable under real production conditions.
His patent portfolio, including US20250371085A1 on enterprise-aware data security posture management and provisional filings on AI agent security assurance and adaptive enterprise intelligence, reflects how that engineering approach has developed into concrete technical inventions addressing the governance challenges organizations are currently working to solve.
“Responsible AI is an engineering discipline,” Singhal said. “It requires the same rigor, the same attention to failure modes, and the same commitment to production reliability that any other critical system demands. The principles are easy to state. Making them hold up at scale, in production, under real conditions, is where the actual work is.”