Paragon Reflex

A product line of Univault Technologies LLC

Paragon Reflex is a reflex layer and a planning LLM for autonomous machines: drones, robots and agents that must keep written rules, hold when they cannot tell, and leave a record of every decision. It is built and sold by Univault Technologies LLC, an AI research company in Salt Lake City, Utah. It is not related to Paragon 28 or its R3FLEX orthopaedic products.

On your receipt

Univault Technologies LLC is the merchant of record for every Paragon Reflex purchase: the dev kit, the prepaid verdicts, and any invoiced work. The name on your card statement and your receipt is ours: Univault Technologies LLC, 1416 S 500 E, Salt Lake City, UT 84105, UEI WJG2GXE99VM5. Questions about a charge go to [email protected] or to the support address on the receipt.

Where the work lives

For an integrator: the facts of the layer

  • What it is. A second detection channel on the same sensor input as your perception stack. It shares no model, no training pipeline and no inference runtime with it, and flags the frames where the two disagree.
  • What it is not, today. At its current false-alarm rate it is a monitoring, arming and event-capture channel. It is not a brake authority, and no assessor has given an opinion on it yet.
  • Decision cost. p50 26.86 microseconds, p99 31 microseconds, on an NVIDIA Jetson Orin Nano at 25 W over 10,000 queries. Under one percent of a detector already running on the same machine.
  • Footprint. The pattern store is 8,192 bytes regardless of how many patterns it holds; one learned pattern moves as a 4,096-byte payload; the released model is 941,316 parameters and its digest does not change when the store does.
  • Platform. Measured on JetPack 6.2 (L4T R36.4.3, Cortex-A78AE). The port to a lockstep safety microcontroller is in build and has not been run.
  • How to buy. The dev kit and prepaid verdicts by card at paragonreflex.com. Operator and program seats by invoice: purchase orders accepted, W-9 and ACH on request, never a card link. No government-furnished equipment or data is required for any engagement.

What is measured, on the record

The deciding parameter set of the reflex is byte-identical after 1,000 real field enrollments, checked by SHA-256 and reproducible by anyone; a decision takes about 27 microseconds on an NVIDIA Jetson Orin Nano over 10,000 queries. These are robotics measurements. No trading number exists, and none is claimed anywhere.