Responsible AI

1. Our Approach to AI

Artificial intelligence is central to how MentorNeko matches mentors with mentees. We believe AI should augment human decision-making, not replace it. By default, AI-generated match suggestions require administrator review before participants are connected. This page describes how we build, deploy, and govern AI features responsibly.

2. How AI Is Used

MentorNeko uses AI in the following areas:

• Match scoring and ranking based on profile compatibility across configurable dimensions
• Match rationale generation that explains why two people were paired
• Text embeddings for profile similarity search and pre-filtering
• Networking introduction group formation to create meaningful small-group connections

AI does not make final decisions about people on its own. By default, match suggestions are presented to program administrators, who have full authority to approve, reject, or override any recommendation. Organizations may optionally enable automatic approval for AI-generated matches at their discretion.

3. Privacy by Design

• Profiles are anonymized before AI processing. Names become generic labels (“the Mentee,” “Mentor A”) and email addresses and phone numbers are stripped from bios.
• Zero Data Retention (ZDR) is enforced on all AI provider API calls through Vercel AI Gateway. No prompts or outputs are stored, cached, or used for model training.
• Real names are substituted back into rationales only after the AI model has finished processing.
• All AI prompts explicitly instruct models that “timezone values indicate time zone, not physical location” to prevent geographic inference from timezone data.

4. Bias Mitigation

• Profile dimensions used for matching are configured by each organization's administrators, not predetermined by MentorNeko.
• Sensitive attributes (such as gender) use opt-in, fail-open filters: exclusions apply only when both participants have set explicit preferences. Missing data is never treated as a signal.
• AI scoring operates on professional compatibility dimensions (skills, goals, experience, interests), not protected characteristics.
• By default, match results are reviewed by human administrators before any participant is notified. Organizations that enable automatic approval retain the ability to terminate or override matches at any time.
• Organizations can configure dimension weights to align matching priorities with their program goals.

5. Human Oversight

• By default, AI-generated matches require administrator approval before participants are connected. Organizations may enable automatic approval if they prefer a streamlined workflow.
• Administrators can override any AI recommendation.
• Match rationales are visible to administrators and participants, providing transparency into why a pairing was suggested.
• Organizations retain full control over which profile dimensions are used for matching and their relative importance.

6. Model Selection and Updates

• MentorNeko uses multiple AI models from different providers (Anthropic, Google, OpenAI) accessed through Vercel AI Gateway.
• Model selection is centralized and version-controlled. Changes to AI models are tested before deployment.
• Customers are notified of material changes to AI model providers or processing approaches.
• We evaluate model performance for match quality and monitor for degradation.

7. Data Governance

• No customer data is used to train, fine-tune, or improve any third-party AI model.
• Our Data Processing Agreement covers AI processing activities. See our DPA for details.

8. Contact

Questions about our AI practices? Contact us.