Openeye

A persistent memory layer for AI agents verifying assembly steps and field procedures through AR/XR cameras. You bring your own vision model (Claude, GPT-4o, Gemini, local Ollama), it turns camera frames into scene descriptions, then runs an agent with tools for step verification, session recall, and procedural skill storage. Every verification becomes a pass/fail/uncertain outcome that gets packaged into ShareGPT trajectories and DPO preference pairs for fine-tuning. Ships with FTS5 search across frames and prior sessions, a FastAPI sidecar that auto-spawns, and reference adapters for both cloud and local vision models. Most traction in manufacturing QC and technician training where mistakes during onboarding are cheap and regulatory overhead is low.