Privacy-Preserving Multilingual Educational AI

Research Problem

How can an AI system route educational queries to optimal inference providers based on real-time data sensitivity classification, while ensuring student PII never leaves the school’s jurisdictional boundary?

Technological Innovation

We are developing a multi-provider AI routing engine that classifies each request across four sensitivity levels — public, internal, PII, and restricted — and routes it to the appropriate infrastructure: self-hosted open-weight models on dedicated EU GPU infrastructure for sensitive operations, or external providers for non-sensitive tasks. A shadow-logging pipeline records parallel provider responses asynchronously, building a proprietary education-domain training corpus for future fine-tuned models.

Why This Is Hard

Sensitivity classification must complete in under 50 milliseconds to avoid degrading the user experience, which requires a lightweight local classifier trained on an education-specific data taxonomy — no such taxonomy exists as a standard. Self-hosted LLM inference at education scale (30+ concurrent classroom sessions) faces GPU memory management, quantization trade-offs, and throughput scaling that are active research problems.

Success Criteria

Greater than 95% sensitivity-detection accuracy at under 50ms latency, validated through an automated test suite with human-reviewed edge cases. Conservative defaults route to self-hosted inference whenever classification is uncertain — a technical privacy guarantee rather than a policy promise.