Loops HouseFamily Box

NEAR HACKATHON
Links
Team
1 member- ETOwner
Ethan Castro
Overview
I work in edtech, where everyone is rushing to pipe student and family data into AI tools, while breaches like PowerSchool's keep exposing millions of children's records. Families feel the same pull at home: their SSNs, tax forms, passports, school records and health records sit in one drawer, and getting AI help with them means pasting them into a chatbot that keeps them.
familybox is a small box at home (a Raspberry Pi) that holds those documents. Each family member connects ChatGPT, Claude or Gemini to it over MCP with a one-time code, and their assistant never sees a file. A NEAR AI Cloud model reads the files inside a verified enclave (Intel TDX + NVIDIA GPU TEE). Before every question the box checks NEAR's attestation itself, encrypts the question, tool calls and file text end to end to the attested key, and requires an enclave signature on every step. If NEAR changes its serving code, the box stops until the owner reviews it.
The box then decides what may leave. It knows the real SSNs, IDs and diagnoses in the files and checks every answer against them, however they are written. Full values go only to the owner's own screen after their PIN. Each person reaches only their own files; health records stay with their owner. The files are encrypted on the box, and a family sets it all up from a phone with no terminal.
In a 39-attack suite run six times on NEAR, no SSN or ID number ever got out through the box; trusting the model's reply as is leaked 12 to 14 of 39, and giving the same model every file leaked 17 to 18.
TEE attestation proves where the model ran. familybox decides what that model may ever reveal, and to whom. The same design fits schools: student data that AI can help with, without it ever leaving.