Loops HouseQuery Sharding

NEAR HACKATHON
Links
Team
1 member- NEOwner
Nerla Jean-Louis
Overview
Users want the quality of frontier closed-source models, but sending them a sensitive question exposes it to the model operator. Query Sharding lets users get closed-model answers without revealing the sensitive details of their question.
A model running in NEAR AI’s private TEE first replaces sensitive entities (people, companies, amounts, hostnames) with placeholders, then splits the question into generic sub-questions that don’t reveal the overall situation. Before anything leaves the TEE, each sub-question passes a leak check: rule-based filters plus LLM judges, including an adversary that tries to reconstruct the original question from all the fragments together. Safe sub-questions are sent as separate requests through NEAR’s incognito mode [to different closed-source providers]. The TEE model then combines the answers and restores the original details, so the full question and the final answer exist only inside the TEE. If the question can’t be split safely, the router fails closed and answers with the TEE model alone.
The approach builds on ConfusionPrompt (Mai et al.), which showed that decomposing prompts lets users benefit from stronger online models while protecting privacy. We add TEE-hosted sanitization and recombination, an explicit leak check with fail-closed fallback, and routing across multiple providers.