AI Trainer - Agent Evaluation Infrastructure (MCP)

At Mindrift , innovation meets opportunity. We believe in using the power of collective human intelligence to ethically shape the future of AI.

The Mindrift platform, launched and powered by Toloka , connects domain experts with cutting-edge AI projects from innovative tech clients. Our mission is to unlock the potential of GenAI by tapping into real-world expertise from across the globe.

We’re on the hunt for hands-on Python engineers for a new project focused on developing Model Context Protocol (MCP) servers and internal tools for running and evaluating agent behavior. You’ll implement base methods for agent action verification, integrate with internal and client infrastructures, and help fill tooling gaps across the team.

Who we're looking for :

Calling all security researchers, engineers, and penetration testers with a strong foundation in problem-solving, offensive security, and AI-related risk assessment.

We’re looking for someone who can bring a hands-on approach to technical challenges, whether breaking into systems to expose weaknesses or building secure tools and processes. We value contributors with a passion for continuous learning, experimentation, and adaptability.

What you’ll be doing :

• Developing and maintaining MCP-compatible evaluation servers
• Implementing logic to check agent actions against scenario definitions
• Creating or extending tools that writers and QAs use to test agents
• Working closely with infrastructure engineers to ensure compatibility
• Occasionally helping with test writing or debug sessions when needed

Although we’re only looking for experts for this current project, contributors with consistent high-quality submissions may receive an invitation for ongoing collaboration across future projects.

Requirements

We also value applicants who have :

Benefits