Intro: A new era for online trust

Founded in 2023, Musubi is on a mission to bring a new level of trust and safety to the digital world. Led by co-founders Tom Quisel (CEO) and Filip Jankovic (Chief AI Officer) Musubi developed a novel Trust & Safety toolkit. Their solution seamlessly blends AI policy enforcement with an adaptive, agentic user-level moderation engine, designed to translate custom content policies into scalable AI that instantly detects violations and continuously learns from moderator decisions. With $8 million in pre-seed and seed funding, Musubi is committed to helping social media platforms, dating apps, marketplaces, AI companies, and gaming companies—any business with user-generated content—to keep their communities safe and thriving.

Challenge: Overcoming black-box moderation

Before Musubi, trust & safety leaders faced a significant hurdle: existing content moderation at scale often relied on static, "black-box" machine learning models. These models were either trained with fixed policy definitions from a third-party vendor, making them rigid, or required immense engineering effort from internal teams to retrain for evolving policy needs. This meant platforms struggled to adapt quickly to new and emerging harms, leaving communities vulnerable and moderation teams overwhelmed. The core problem, as Musubi observed, was that "trust & safety leaders trying to moderate user-generated content at scale were stuck with black-box machine learning models which were either trained with fixed policy definitions (from a vendor) or took huge effort for their engineering teams to retrain."

Solution: Dynamic moderation, powered by Google AI and Gemini

Musubi harnessed the power of Google AI, particularly Gemini models, to create a dynamic and highly responsive moderation system. They use Gemini models for sophisticated user-generated content labeling and classification, allowing trust & safety teams to effectively remove harmful and illegal content and protect their users. Their innovative PolicyAI tool sends content and custom prompts—incorporating community guidelines and specific instructions—to Gemini. The output is a precise classification and label, along with a severity score and reasoning. This approach empowers policy experts to define exactly what their policy should be and update it in minutes based on new harms. "By using Gemini, we allow policy experts to define exactly what their policy should be, based on their unique community's needs, and update their policies based on new and emerging harms in minutes," Musubi explains. They've found that Gemini's balance of accuracy, latency, and cost, across models like Gemini 3.1 and Flash, is unmatched by other solutions they've tested.

Results: Unprecedented precision and speed

The integration of Gemini has brought remarkable precision and speed to Musubi's customers. While metrics vary by platform, Musubi consistently observes F1 scores of 90%+ for content classification, frequently outperforming human moderation teams. In three recent pilots, they achieved 90-95% F1 scores on content classification and labeling. Indeed, in some cases, they’ve seen 98% F1, especially with thinking models. This has allowed customers to enforce their own unique policies at scale, rather than relying on generic models. "Our customers are thrilled that they are finally able to enforce their own, custom, unique policies at scale, instead of relying on black-box ML models with fixed ideas of what 'harm' is," Musubi reports. One major social platform even used PolicyAI + Gemini LLMs to launch a new feature at scale, confident that user-generated content would remain safe and appropriate from the start.

What's next: Scaling and innovating with Gemini
Musubi is rapidly scaling, with much of their continued development built on Gemini integrations. They’ve recently added new features for LLM oversight and governance, including automated prompt optimization and diagnostics. They've also expanded multimodal capabilities, recognizing that "their customers are very excited about multimodal capabilities, as this is something that has been difficult to moderate in the past." Musubi sees this as an incredibly exciting time for trust and safety, as AI allows platforms to "respond faster, apply custom rules and policies, and adapt in real-time in a way that they weren't able to do before." For founders on the fence about using Gemini, Musubi's advice is simple: "Gemini makes it easy to try things out, and their self-serve process is great, so we would recommend just checking it out and seeing how it performs. The ecosystem is incredibly healthy, making it a safe long-term choice." Musubi continues to push the boundaries of what's possible, ensuring a safer, more trusted online experience for their 300 million end-users.