Quarks improves user experiences with Gemma 2 and Gemma 3 With the help of Gemma 2 and Gemma 3, Quarks delivers safer and more engaging experiences on their dating and connection apps. Quarks is a business with Ukrainian roots that builds apps like Kismia, Catchyy, and Affemity for online dating and social connection. Used around the world, these apps focus on helping users build deeper, more meaningful connections. To achieve this, Quarks places a high standard on trust, safety, and personalization.To further meet its standards and streamline workflows, Quarks used both Gemma 2 and Gemma 3 to improve the messaging experience, help moderate user-generated content, tag images more effectively, and process user feedback across apps.The challengeMessaging is one of the core—and most challenging—elements of a dating or connections app. The Quarks team observed that many users struggle to not only initiate conversations but also to keep conversations going with other users. And if users aren’t messaging other users or keeping each other engaged, they may lose interest and leave the app entirely. The team introduced a suggested messages and replies feature using bespoke content, but the suggestions were often too generic and broad, failing to cater to all of the possible combinations of conversational contexts and user preferences.Another challenge is content moderation. With over 20 million new pieces of user-generated content added to its apps each month, Quarks needed a solution to reduce the burden on its content moderation team, while also effectively detecting things like scams, fraud, and policy violations. “Manual moderation was a bottleneck,” said Oleksii Avilov, AI Platform Team Lead at Quarks. “ [Gemma] helps us moderate 20+ million content items per month, assist in conversations, tag millions of photos, and analyze feedback — not just faster, but smarter. With AI, we build systems that learn, adapt, and improve, just like the humans using them. Oleksii Avilov, AI Platform Team Lead at Quarks Manually reviewing images also proved challenging. Over 4 million images are uploaded to Quarks’ apps monthly, again necessitating a large image moderation team and an extended review period that could impact safety and privacy for users. Additionally, the team needed help not only moderating images but also tagging them with relevant metadata to improve searchability and curation for users.Lastly, the team faced a high volume of unstructured feedback. Per month, Quarks receives over 100,000 pieces of text-based reviews, complaints, and feedback from users. Parsing through that data to find common themes and determine which topics are most important to address can also be time consuming, making the Quarks’ team slower to respond to issues. Chart representing gains made in profile description reviews with Gemma. The solutionQuarks tested multiple open models to find a solution, and found that Gemma 2 and Gemma 3 27B offered the best performance, accuracy, reasoning ability, and deployability. “The 27B model consistently performed better in tasks that require a deeper understanding of context, including scam detection, content moderation, and message generation,” said Avilov.An added benefit of Gemma 27B is that its strong performance relative to its size means it can run on a single GPU (e.g. NVIDIA RTX 4090 or A100 80 GB), greatly simplifying the deployment process and keeping infrastructure costs down. This scale means on-premise inference is possible, keeping sensitive user data private. Chart representing improvements made in review time with Gemma. To improve messaging on Quarks’ dating and connection apps, the team built a real-time conversation assistant that uses Gemma 2 and Gemma 3. This assistant is able to create context-aware message suggestions that are relevant and tailored to chats between users to keep them engaged longer and increase retention.For content moderation, the team built a hybrid moderation engine that uses Gemma 2 and Gemma 3 to identify and flag inappropriate, scam-like, or fraudulent written content for review by the moderation team with over 95% accuracy. For image moderation, the team employs Gemma 3 to support human moderators in reviewing images and applying helpful metadata to improve user experience and safety. These tags can also add another level of security by flagging possible visual policy violations early in the review pipeline. “ Gemma excels at detecting subtle violations, such as obfuscated contact information, scam narratives, or ambiguous adult content, that traditional rule-based systems or lightweight classifiers often miss. Oleksii Avilov, AI Platform Team Lead at Quarks Gemma also helps the team efficiently analyze user feedback to spot common themes like pricing concerns, trust issues, or UX pain points, among others, saving valuable time and enabling faster responses from the product and support teams.All AI-powered tools at Quarks are designed with user privacy and data security in mind. Human oversight is maintained throughout moderation workflows, enabling the team to work more effectively while still meeting high standards of safety for its users.The impactQuarks has observed great results since leveraging Gemma 2 and Gemma 3. The real-time conversation assistant now supports over 1 million messages per month and has improved user reply rates and satisfaction, leading to higher retention and engagement metrics.Gemma has also helped Quarks dramatically improve its efficiency by pre-filtering written content before the human review process. Gemma's pre-filtering and the improved efficiency in the review process led to fewer scam reports from users and a lower overall review volume for human moderators. The team has also experienced improvements to the image review and tagging process, reducing their workload and improving searchability and curation for users.These gains in efficiency and accuracy have helped the team at Quarks simplify their processes and deliver better user experiences more quickly, leading to increased user satisfaction and higher enrollment numbers.What’s nextThe Quarks team plans to integrate AI further into their workflows. “We’re actively expanding our use of LLMs beyond core moderation and messaging assistance into more personalized, real-time user experiences,” said Avilov. The team wants to create a private AI companion that can guide, support, and engage users based on their preferences and behavior, while ensuring safety and privacy for a more engaging experience.To further simplify their review and moderation processes, the team is looking into multimodal integration to combine its text and image review pipelines to more holistically identify nuanced policy violations, fake profiles, and inappropriate content.For Gemma specifically, the team is currently testing to see how ShieldGemma 2 can help reinforce content safety. They are also considering edge deployments, testing Gemma 3 12B for environments where lower latency and compute constraints are critical.Avilov summarized the benefits Gemma has brought to the Quarks team, and what it enables them to offer their users: “It gives us the tools to better understand people at scale, to personalize their experience, to protect them from harm, and to help them express themselves—all without losing that sense of authenticity.”