HQZN / TECHNOLOGY & R&D

Technology & R&D

From model capability to a system that can be verified.

We develop data interfaces, software components and validation tools for practical business needs, combining established computing platforms with open-source technologies.

LOCAL INFERENCE / EXPERIMENTAL VALIDATIONAugust 2026

One task. Different edge configurations.

Comparing local open-model inference across platforms and quantization configurations to inform on-device application development.

Development platforms: NVIDIA Jetson AGX Thor; NVIDIA DGX Spark for development and cross-validation.
Tasks: Code, mathematics and text generation, with fixed model and software versions, repeated runs and output checks.
Method: The common protocol includes 3 warm-up runs and 5 measured runs per test, with numerical probes and fixed-corpus quality evaluation where supported.
Output: Platform- and configuration-level compatibility and performance records to inform workload-specific inference choices.

Validation scopeLimited to the tested versions and single-request experiments. Some configurations lack full compatibility or quality validation. Concurrency, long-context and end-to-end vision were not validated in this study; it does not establish fleet-wide production readiness.

These third-party names are trademarks or registered trademarks of NVIDIA Corporation and identify platforms used in our research only.

Correctness before performance

Evaluate tasks, datasets and baselines together. Preserve experimental conditions as well as unsuccessful results.

Separate generation from verification

Check code and conclusions with independent validation, tests and human review, then turn lessons into reusable tools.

Keep execution accountable

Models support analysis and recommendations. Business rules constrain actions, and critical operations retain human approval.

RESEARCH INFRASTRUCTURE

Technology choices and their uses

Technology choices and their uses
Model computing & evaluation	NVIDIA Hopper architecture / CUDA	Hopper-based computing for model deployment, domain training experiments and business-oriented evaluation.
Edge AI development	NVIDIA Jetson Thor / TensorRT Edge-LLM	Local inference, quantization configurations and edge interaction prototypes.
Development & cross-validation	NVIDIA DGX Spark	Inference framework comparisons and cross-platform issue verification.
Digital twin research	OpenUSD / NVIDIA Omniverse Kit / RTX	Asset conversion, telemetry mapping and scene-rendering prototypes.
Embedded devices	RK3568	Device integration, on-board small-model experiments and basic interaction.

Names identify technologies used in our R&D, without implying product certification, an authorized partnership or third-party endorsement.

NVIDIA Corporation owns the NVIDIA, CUDA, DGX, Hopper, Jetson, Omniverse, RTX and TensorRT trademarks and/or registered marks in the U.S. and other jurisdictions. All other third-party names and marks belong to their respective owners.

What we develop

Data adapters and business APIs; tariff and replay tools; model-serving gateway components; asset conversion and state mapping; evaluation and development automation tools. Foundation models, open-source frameworks and hardware are attributed to their respective creators.

Continuous development. Durable knowledge.

We explore an R&D workflow of human-defined goals, AI-assisted execution and independent verification, building reusable data, code, evaluations and knowledge through each iteration.

LET’S BUILD WHAT’S NEXT

Let's talk equipment, data and operations.

Discuss device integration, operations software, on-device inference or digital twins with our team.