Brandyn Leonard — Research Lead, ML Architect, Builder

About

Lead architect building original systems from concept to execution.

I design the architectures. LOLM’s hybrid Transformer-SSM dual-stream system, the traffic network operating-band framework, the g(K) = 4K(1−K) gating function in EGC — these are mine. I built them because no existing tool solved the problem the right way.

Co-founded Qira LLC with my brother Bryan Leonard. Self-taught — no university, no lab, no venture capital. Currently scaling models to 1.57B parameters on NVIDIA H200 and Google TPU v4-8 infrastructure while running a live traffic intelligence system and an empirical consciousness study.

U.S. Provisional Patent #64/002,166 filed March 10, 2026.

Open to: Research partnerships, compute grants, municipal technology pilots, and ML teams tackling real problems at real scale.

PyTorch JAX/TPU Python CUDA NumPy SciPy Transformers SSM Kuramoto FastAPI React TypeScript

Work

Flagship — Patent Filed

LOLM — Latent Order Language Model

A hybrid Transformer-SSM architecture that separates language into surface and latent streams. The 29% minority SSM path turns out to be 8,645x more essential than the 71% majority Transformer path — dependency inversion. 7 complementary training losses, 5-stream fan-out architecture.

43% faster convergence at 300M · 15% lower perplexity at 1.57B · H200 + TPU v4-8

Three original theoretical frameworks guide the architecture: a field-theoretic gate mechanism, a multi-scale phase selection system, and a latent generative process model. Each maps to a specific component and training objective.

Patent #64/002,166

GitHub → Project site → Qira HQ →

Original Statistical Framework

Traffic Network Intelligence

An original framework for detecting robust coordination regimes in networked systems. The core innovation: operating bands (ridges) instead of fragile point optima. DON translates these ridges into real traffic signal adjustments with safety clamping.

207 sensors · 34K+ time points · 1,191 edges · Max-T Z* > 4.8, p < 0.01

Ridge fraction 95–97% · Multiplex scale-free cross-layers strengthen stability · Validated on METR-LA benchmark

Validated

Technical dashboard → Public corridor monitor →

Live Empirical Study

EGC — Expression-Gated Consciousness

A mathematical framework measuring how emotional knowledge gates conscious expression. The gating function is mine: g(K) = 4K(1−K). Live study with 44+ participants, three confirmed response types, and a stable correlation from N=14 forward.

g(K) = 4K(1−K) — Brandyn Leonard

N=44+ · Pearson r = 0.311 · 3 response types confirmed

Live

Take the study → The Gate — research page → Preprint on Zenodo →

Phoenix Traffic Intelligence

Real-time traffic intelligence monitoring 8 Phoenix freeway corridors 24/7. Live AZ-511 data, TomTom live speeds, cascade prediction, FHWA delay cost analysis, and AI-powered crew dispatch recommendations every 2 minutes.

408K+ corridor snapshots · 51K+ network snapshots · ~$3.99M projected annual savings

Live 24/7

Public corridor monitor → Technical dashboard →

Key Numbers

Measured. Validated. Reproducible.

8,645

x Dependency Inversion

The 29% minority SSM path in LOLM is 8,645 times more essential than the 71% majority Transformer path. This was not predicted. It was discovered.

43%

Faster Convergence

LOLM’s hybrid architecture converges 43% faster than comparable single-stream models at the 300M parameter scale. Same data, same compute, fewer steps.

15%

Lower Perplexity

At 1.57B parameters, LOLM achieves 15% lower perplexity than baseline. The dual-stream separation is working — the latent path captures what the surface path cannot.

207

METR-LA Sensors Validated

Validated on the full METR-LA benchmark: 207 sensors, 34,272 time points, 1,191 edges. Max-T corrected significance at p < 0.01.

1.57B

Parameters Scaled

LOLM tested at 20.5M, 149M, 304M, and 1.57B parameters. Each scale validated the architecture. The dependency inversion holds across all of them.

0.311

Pearson r (EGC)

A stable correlation between expression and comfort that has held from N=14 to N=44+. Three distinct response types confirmed in the data.

97%

Ridge Fraction

95–97% of the operating space in validated networks shows ridge behavior, not fragile point optima. The system is robust by structure, not by accident.

Institutional Backing

No university. No lab. No venture capital. U.S. Provisional Patent filed. Every system self-funded and self-built from a home office in Maricopa, Arizona.

Key Discoveries

Dependency Inversion

In LOLM, the minority path (29% SSM) turns out to be 8,645x more essential than the majority path (71% Transformer). The architecture that does less structural work carries almost all the semantic weight. This inverts the assumption that majority components dominate.

8,645x essentiality ratio · Consistent across 4 model scales

Gradient Isolation

LOLM’s 7 complementary losses create natural gradient isolation between streams. Each loss function trains a specific capability without interfering with others. This is what enables the 43% convergence speedup — no loss is fighting another.

7 complementary losses · 5-stream fan-out · Zero-conflict training

Two-Phase Gate Trajectory

In the EGC data, subjects don’t move linearly from low expression to high expression. They follow a two-phase trajectory governed by g(K) = 4K(1−K) — rising gate openness until K=0.5, then falling. The gate has a maximum, not a monotonic climb.

g(K) = 4K(1−K) · Peak at K = 0.5 · 3 response archetypes

Operating Ridges, Not Point Optima

Our framework shows that real networked systems don’t have fragile optimal points. They have operating ridges — broad bands where the system is stable and performant. 95–97% of the validated operating space is ridge. Optimization means finding the band, not the point.

95–97% ridge fraction · METR-LA validated · Safety-clamped DON output

Philosophy

I build the thing. Not the pitch for the thing. Not the roadmap for the thing. The actual thing. If the architecture doesn’t work at scale, the idea was wrong. Fix the idea.

The best systems aren’t fragile at their optimum. They have ridges — broad operating bands where everything works. I design for ridges. In architectures, in frameworks, in how I work.

Self-taught is not a limitation. It’s a design constraint. No institutional inertia. No committee approval. No inherited assumptions. Every decision I make, I made because the evidence said to.

Contact

Architectures, research, and hard problems.

Email Brandynleonardtech@gmail.com GitHub github.com/TheArtOfSound LinkedIn linkedin.com/in/brandyn-leonard Preprint zenodo.org/records/19242315 Traffic adot.autohustle.online Traffic Intelligence nfet.autohustle.online The Gate theartofsound.github.io/thegate

For ML Researchers & Labs

LOLM’s dependency inversion finding challenges the assumption that majority architecture components dominate. The 8,645x essentiality ratio is reproducible. The patent is filed.

View the architecture on GitHub →

For Transportation & Municipal Leaders

The traffic coordination framework is validated on 207 sensors with statistical rigor. The Phoenix traffic system runs 24/7 on live data with projected savings of ~$3.99M annually per 100K vehicles.

See the live corridor dashboard →

For Consciousness & Psychology Researchers

The EGC framework proposes that expression gates consciousness through a measurable quadratic function. The study is live, the preprint is on Zenodo, and we’re looking for collaborators to scale it.

Take the study or read the preprint →

For Compute & Cloud Providers

We are training at 1.57B parameters on H200 + TPU v4-8 and need to go further. The architecture works. The patent is filed. We need compute to prove what it can do at frontier scale.

View the LOLM project site →