Blindsample

BlindSample is a private data evaluation platform that helps buyers assess a dataset before buying it without revealing the seller's records. A buyer creates an evaluation and asks plain-language questions about completeness, consistency, accuracy, relevance, or quality. BlindSample then creates private buyer and seller links. The seller opens their link and uploads a small CSV sample. The sample and every buyer question are sent together to 0G Private Computer for confidential AI inference. The seller's rows are not displayed to the buyer or stored as part of the published result. Instead, the buyer receives one score per question, with a clear explanation, confidence level, scoring anchors, and safe evidence. BlindSample also exposes verification metadata, including the 0G model, provider, request ID, token usage, and TEE proof. If the 0G request or verification fails, the platform publishes no score. This makes private dataset discovery safer, clearer, and easier to audit for teams.

BlindSample is built with Next.js, TypeScript, React, Supabase, and Vercel. Supabase stores evaluation metadata, buyer questions, access permissions, and final results—but never the seller’s raw CSV sample. Private capability links give buyers and sellers access to their separate workflows. When a seller submits a CSV, a Vercel server function validates and holds the sample temporarily in memory. It sends the sample and all buyer questions in one request through the 0G Router to 0G Private Computer. The model runs inside a trusted execution environment and returns one structured score per question, including an explanation, confidence level, scoring anchors, and privacy-safe evidence. BlindSample also records the 0G model, provider, request ID, token usage, and TEE verification metadata as proof of execution. A notable design choice is that there is no local scoring or AI fallback: if 0G inference or verification fails, BlindSample publishes no score. We also allow buyers to ask ordinary plain-language questions instead of forcing them to configure columns, schemas, or scoring rules. To control cost and privacy exposure, each evaluation uses a bounded CSV sample of up to 50 records.