PARAGON REFLEX

Microsecond-class reflex intelligence for embodied AI.

Make the self-driving
dream possible.

Production AV (autonomous vehicle) stacks miss pedestrians at highway speed because their reflex loop runs in tens of milliseconds. Our reflex kernel runs in 4.23 microseconds (measured on an A100 GPU).

A deterministic reflex layer that sits beneath the perception stack — designed to catch what the network misses, in time to brake.

Signed & reproducible
Reflex recovers what a 2024 SOTA (state-of-the-art) detector misses
~4 µs
p50 Latency
86.77%
Network-miss catch
84.9-88.5%
95% CI
Argoverse 2 · LION-Mamba (2024 SOTA)
· 6 cities · n=1,368 · 2% FPR · Ed25519-signed

A product line of Univault Technologies. Currently in active discussions with Tier-1 automotive suppliers. Series A allocation closing.

~4 µs p50 | 86.77% Network-miss catch (signed) | Argoverse 2 + nuScenes | Cross-NN validated | ASIL-D roadmap | Patent-protected

Why this matters

The dream of self-driving
has a timing problem.

Modern AV perception is brilliant at understanding scenes, and slow at the moment when speed is the only thing that matters. A car at 25 m/s travels half a meter in 20 milliseconds. The pedestrian is already in the lane.

Today's reflex gap
  • Perception loop runs at frame rate
    Tens of milliseconds per inference. Fine for planning, slow for emergencies.
  • Edge cases miss silently
    Out-of-distribution events bypass the network's confidence layer entirely.
  • Cross-region performance drops
    A model trained in Boston degrades sharply when deployed in Singapore.
  • No deterministic safety floor
    ASIL-D (Automotive Safety Integrity Level D) certification of a learned-only stack is an open problem.
With a reflex layer
  • Microsecond-class response
    4.23 µs p50 reflex-kernel latency on an A100 GPU, orders of magnitude faster than a learned perception loop.
  • Catches what the network misses
    A second, deterministic decision path for the patterns the network was never trained on.
  • Holds across regions
    Cross-geographic gap reduction measured on nuScenes, Boston to Singapore.
  • Auditable safety floor
    Deterministic by construction. Track to ASIL-D as a Type-2 SEooC (Safety Element out of Context).

Cross-architecture validation

Same reflex layer.
Three real detectors.
A recovery margin on each.

We evaluated the same reflex layer against three independently developed, real perception detectors on public AV benchmarks. It recovers a meaningful share of the hazards each one misses. The catch rate tracks each detector’s blind spots — 86.77% on LION-Mamba, 64.6% on CenterPoint, 59.2% on BEVFusion. The point is that the reflex layer adds a recovery margin on top of every detector we tested, not just one.

CenterPoint · real detector

64.6%

Network-miss catch rate

nuScenes · n=127 network misses · evaluated on Modal A100-80GB. Signed claim.

BEVFusion · real detector

59.2%

Network-miss catch rate

nuScenes · n=238 network misses · same reflex configuration · Modal A100-80GB. Signed claim.

Why this matters for the OEM (original equipment manufacturer). We tested the same reflex layer against three independent real detectors — LION-Mamba, CenterPoint, and BEVFusion. It recovers a meaningful share of the hazards each one misses (59–87%, tracking each detector’s blind spots). A safety layer that only helps one perception stack is a coupled dependency; one that adds a recovery margin on top of whichever detector the OEM chose is a portable safety floor — because the OEM gets to keep choosing.

What we’ve validated

Three results that move
the safety conversation.

Public benchmark, public dataset, production cloud GPU evaluation — the results we walk Tier-1 partners through line by line.

01 — LATENCY

4.23 µs

p50 reflex latency

Median reflex-kernel latency, measured on Modal A100-80GB. Orders of magnitude faster than a learned perception inference loop. Automotive-runtime port is the next funded milestone.

02 — NETWORK-MISS CATCH

86.77%

of a 2024 SOTA detector’s misses · signed

Reflex recovers 86.77% of the hazards LION-Mamba (2024 SOTA) misses across 6 Argoverse 2 cities (full 0–30m forward cone, 2% FPR (false-positive rate), n=1,368, 95% CI (confidence interval) 84.9–88.5%). Ed25519-signed, reproducible offline.

03 — GENERALIZATION

3 detectors

Cross-architecture, cross-dataset

Validated against three real detectors (LION-Mamba, CenterPoint, BEVFusion) across Argoverse 2 (6 cities) and nuScenes (Boston, Singapore). Catch rate 59–87%, by detector.

Modeled scenario (not a signed claim): paired with a production-grade base detector at 85% TPR (true-positive rate) / 2% FPR, the cascade reaches 94.5% TPR at 4.6% FPR, combining measured Reflex with a modeled production NN (neural network). Per-region tables, stopping-distance numbers, and the full evaluation walk-through are part of our Tier-1 partner conversations.

Watch the demo →

Open roles

Help us put a reflex layer
under every embodied AI.

We’re hiring deliberately. Every seat below either ships the AV product or unlocks a Tier-1 conversation. If you’ve worked on production automotive safety, we want to hear from you.

Talk to us if you ship
autonomy at scale.

Tier-1 supplier, OEM, robotics platform, eVTOL (electric vertical take-off and landing) program — if your roadmap depends on a microsecond-class safety floor, we should be talking.