Loops HouseEyewitness

NEAR HACKATHON
Links
Team
1 member- SEOwner
Sergei Birtanov
Overview
Eyewitness is paid human-in-the-loop for AI agents: when an agent can't continue because it needs to see something in the physical world, it pauses, pays a real person with a phone to look, checks the evidence and resumes. Example: an inventory agent can't read a label, so it commissions a close-up from the same person and finishes its report.
Through our MCP tool, Claude posts a task; the reward plus the task's storage cost go into our NEAR contract, which generates a 4-digit challenge code from the block's random seed. The worker photographs the scene with the code on paper. A local vision model on the hub must read the exact code before payment (it is never told the code). The hub records the photo's sha256 on-chain, the contract pays the worker seconds later, and only then is the photo released. The agent verifies the hash against the task it requested. No delivery before the deadline: anyone can refund the agent.
Honest scope: results are human-submitted, hub-accepted observations; the code is a replay deterrent, not proof a scene is real; the hub is trusted in v1. We had the project strictly reviewed and fixed what it found (exact code match, task-bound verification, per-worker credentials, private-until-paid photos, requester-paid storage).
Built in the window: Rust near-sdk contract (16 unit tests, 12 regression checks), Node hub with Ollama/Qwen2.5-VL check, phone page with the Blink mascot, dashboard, website with live chain stats, MCP server + CLI, verification log of real testnet runs. Testnet: eyewitness-live.testnet. Website with live on-chain stats: https://birserg.github.io/eyewitness/ . Apache-2.0.