From AI Assistants and Photo Enhancement to Moderation and Safety: Integrating Gemini into the Quarks Product Ecosystem

Quarks operates in sectors where interaction quality, content safety, and speed to market directly drive growth. Consequently, the team adopted AI not as a peripheral experiment, but as a core tool to improve critical stages of the user journey.

Before making the broader transition to Gemini Enterprise Agent Platform, the Quarks team had already been actively testing various AI approaches.

The Challenge: Building a Stable AI Infrastructure for Product Scenarios

Initially, several scenarios were built using Gemma 2 and Gemma 3, deploying models on rented GPU resources and later through external AI providers. While this offered flexibility, the growing number of use cases demonstrated that a production-grade AI system required a higher standard of stability, quality, scalability, and control.

The team required a platform capable of:

• Running multiple parallel AI scenarios in production;
• Scaling effectively while maintaining quality and uptime;
• Handling text, images, and multimodal tasks;
• Automating moderation and safety processes;
• Rapidly testing new hypotheses;
• Maintaining strict control over costs and performance.

Operational stability was a primary challenge. Many AI functions at Quarks are integrated into critical user flows, meaning response quality, latency, and availability directly impact the user experience, conversion rates, and brand trust. Consequently, Quarks approached generative AI not as a one-off experiment, but as a systemic product and operational layer designed for scaling across various scenarios.

Why Gemini by Google Cloud?

After evaluating various options, Quarks chose Gemini as the foundation for its expanding AI initiatives.

The decision was driven by practical business factors:

• High quality for real-world product tasks;
• An optimal balance between cost and output;
• Support for multimodal scenarios;
• Rapid integration into existing services and product logic.

Collaborating with Cloudfresh, a Premier Google Cloud Partner, the team refined their application of Gemini. This included selecting specific models for targeted tasks and optimizing performance, request costs, and architecture to ensure reliability under heavy production loads.

Gemini delivered the most balanced results for their production environment. The team began developing a unified AI layer that now supports multiple critical areas, from user interactions to moderation, safety, and internal operations.

Use Case 1: AI Photo Enhancement with Nano Banana

In dating products, photo quality directly influences first impressions, likes, matches, and the likelihood of starting a conversation.

In reality, many users upload photos with poor lighting, low resolution, or hardware limitations. To address this, Quarks tested an AI Photo Enhancement feature that improves image quality without altering facial features or creating an artificial appearance.

This functionality utilizes Nano Banana—Google’s high-speed model within the Gemini stack designed for image generation and editing.

The objective was not to “re-draw” the user but to find a balance between:

• Improved image clarity;
• Preserving authenticity;
• Photorealistic results;
• Avoiding a “bot-like” profile aesthetic.

This is a vital product nuance: in dating, an overly “perfect” photo can actually decrease trust rather than increase it.

Initial Results

While the feature is still in the testing phase, the team is already seeing positive indicators:

• Approximately 5% of users have already engaged with the feature;
• In the test group, the number of likes sent by women increased by 27%;
• Match rates rose by 7%;
• Conversation rates among men grew by 7.8%.

For dating products, these metrics have practical consequences. Photo quality affects not just visual appeal, but user behavior throughout the funnel.

This case demonstrates that AI doesn’t always need to “dazzle” the user. Often, its most valuable role is removing the barriers that prevent users from fully engaging with the product.

Use Case 2: Gemini for AI Assistants in Relationship and Well-being Products

A major AI focus for Quarks is an assistant platform integrated into various products and user interaction scenarios.

In relationship-focused products, the AI assistant helps users in specific situations: understanding their relationships better, articulating needs, building healthier communication, setting boundaries, or finding the right words during emotional moments.

One example is Affemity, an educational platform for women’s mental health. The AI assistant helps users navigate relationships and communication with men, offering support on emotional clarity, personal boundaries, flirting, and mindful partner interaction.

Results Confirmed by User Behavior

This scenario shows clear behavioral evidence that the AI assistant provides genuine value:

• 60% of users interact with the assistant;
• An average of 20+ interactions per user;
• Positive feedback is 8 times higher than negative reactions.

For Quarks, these are more than just engagement metrics. The assistant is becoming part of the core product experience, helping users understand their feelings and communication dynamics.

The functionality was developed in close collaboration with coaches and psychologists. This ensures the assistant is not just “intelligent” but appropriate for sensitive and emotionally complex situations.

As a result, the product achieves differentiation through the depth of user interaction and trust rather than simply adding features.

The team’s next phase focuses on personalization.

Quarks plans to increase the utility of AI interactions by:

• Improving situational awareness for each user;
• Providing more relevant advice;
• Deepening assistant integration with other app features.

Use Case 3: Gemini for Profile Description Moderation

One of the most established AI use cases at Quarks involves platform safety and user trust.

In dating products, user-generated content drives engagement but also introduces risks like fraud, manipulation, policy violations, and harmful content.

The “About Me” profile section is particularly sensitive. Users often include:

• Subtle indicators of fraud;
• Attempts to bypass platform rules;
• Restricted contact information;
• Inappropriate or unwanted content;
• Other high-risk signals.

Before implementing Large Language Model (LLM) moderation, all “About Me” texts were checked manually. As content volume grew, manual review became a bottleneck.

This led to several operational issues:

• Delayed profile reviews;
• Late detection of risky cases;
• Moderators spending excessive time on safe content instead of critical cases.

Essentially, moderation resources were not being utilized where they were most effective.

The Core Strategy: LLMs as an Intelligent Filter

To solve this, Quarks used LLMs not as a simple “approve/reject” tool, but as an intelligent semantic filter to reduce noise and identify risky content rapidly.

Instead of a binary approach, the system analyzes text meaning and classifies it by risk level and type. This allows for tailored moderation workflows based on the potential problem. Low-risk content is processed automatically, while human moderators focus their attention on cases where subjective judgment is critical.

Key Metrics Demonstrating Value

Following LLM implementation, moderation at Quarks is nearly fully automated and significantly faster:

• ~99% of “About Me” content undergoes AI analysis;
• Profile review coverage increased from 17% to 100%;
• Approximately 80% of profiles are automatically approved in less than 5 seconds;
• Reaction time to fraud and safety issues has dropped to under 1 minute;
• Manual moderation is now concentrated solely on high-risk cases.

This directly improves product quality, as users encounter fraudulent or dangerous content much less frequently.

In this scenario, AI supports rather than replaces the moderation team, allowing them to work with precision at a scale manual efforts couldn’t reach. For Quarks, this is linked to product integrity and the ability to scale the platform without losing control over safety risks.

Use Case 4: Gemini + RAG for Contextual AI Responses

Quarks also uses Gemini for scenarios where response quality depends on access to relevant context in addition to the model itself.

For several AI assistants, the team develops workflows where answers rely on internal or specialized knowledge bases rather than just general model capabilities. To achieve this, Quarks utilizes a Retrieval-Augmented Generation (RAG) approach, which integrates necessary context from large datasets, including documents and PDFs materials, into the response.

This approach helps:

• Increase response relevance;
• Reduce the risk of inaccurate or generic recommendations;
• Provide more useful, context-aware advice;
• Quickly scale AI scenarios that require niche expertise.

This marks a transition from simple generation to contextually enhanced responses that provide higher value to the user.

What’s Next: Internal AI Chat and Corporate Knowledge Base

Quarks’ vision for AI extends beyond user-facing features. The next step is developing an internal AI agent using MCP (Model Context Protocol) to serve as a central access point for company knowledge.

The goal is to provide teams with a tool to:

• Locate information faster;
• Navigate company structure easily;
• Access internal knowledge via natural language;
• Analyze reports and internal data through chat;
• Reduce time spent on repetitive internal queries;
• Maintain full control over data access and detail levels according to security policies.

At the same time, data access in this chat remains restricted and controlled based on user roles and information sensitivity levels. The MCP approach ensures that the AI does not have direct access to raw data, working only with authorized, filtered, or aggregated responses. This maintains a balance between usability and security, minimizing the risks of data leaks or improper use of internal information.

Partnership with Cloudfresh

Throughout this project, Cloudfresh helped Quarks move AI initiatives from the experimental phase into live product environments.

This collaboration allowed Quarks to launch new AI scenarios faster, manage stability and costs effectively, and focus on product value rather than infrastructure hurdles.

Key areas of cooperation included:

• Selecting the best Gemini models for various use cases;
• Managing Gemini Enterprise Agent Platform performance and reliability in production;
• Optimizing AI requests, token usage, and multimodal workflows.