We built a 1KB/node graph system with 0(1) lookup and semantic recall

We’ve been experimenting with a new database architecture for real-time AI, robotics, and agent workloads. The prototype in this 90-second demo shows a persistent graph engine that:

delivers 0.12–0.4 µs hot-path reads

stores 50M durable nodes on an 8GB Jetson Orin Nano

streams a 40–50GB graph from NVMe as if it were RAM

maintains ACID guarantees under power-fail/crash tests

uses a hardware-native concurrency lattice instead of locks or B-trees

The goal wasn’t to optimise an existing database, but to test whether durable sub-microsecond access was even possible without keeping the whole graph in memory.

We’re a very early startup exploring this space and would genuinely appreciate critical feedback, collaboration, and technical pushback from people who’ve worked on database internals, kernels, robotics systems, or high-performance storage.

If you see flaws, edge cases, missing failure modes, or things that “shouldn’t work,” we’d love to hear from you.

Happy to run the demo live for anyone curious, just reply or DM. Also open to free pilots with robotics/defence labs.