Purpose
ZETIC.ai is an edge AI company whose goal is to bring AI into products running on real hardware. Its current focus is robotics and Physical AI, building on its work in mobile and embedded inference. Its services are aimed at teams that need to run models locally on devices and want faster inference speeds, reducing wait times through hardware-aware execution design, improving the performance of existing hardware, and reducing the team's own optimization workload.
Main capabilities
- Hardware-aware inference optimization: Officially described as "execution design for faster inference," that is, designing and optimizing the way models are executed on the target hardware to achieve faster AI responses.
- Measured before-and-after optimization comparison: Provides comparison data between baseline execution and optimized execution on the same hardware to quantify performance changes.
- ROS 2 integration package: The deliverable includes a ROS 2 software package and integration guidance for easy integration into robotic systems.
- Hardware selection comparison: Before purchasing a development board, compare supported candidate hardware options for the user's workload.
- Retain existing backends: The official description states that TensorRT can be retained as a backend while ZETIC optimizes the workload.
How to use
The officially described workflow is:
- The user provides the model + target hardware, as well as performance and quality goals.
- Baseline measurement is performed.
- Execution design is carried out.
- Validation and packaging.
- Delivery of the integration package (including the ROS 2 software package and integration guidance, the optimized execution plan, and before-and-after optimization measurement results).
The official description states that the scope of work and the required validation inputs are confirmed before work begins.
Use cases
- Running AI inference on robotics hardware and wanting ROS 2 integration.
- Improving AI performance on existing development boards and reducing response wait times.
- Comparing how different hardware options fit one's own workload before purchasing a development board.
- Wanting execution optimization while retaining existing backends such as TensorRT.
Officially mentioned supported hardware platforms include: NVIDIA Jetson Thor, Orin, Nano; Qualcomm QCS series, iQ series.
Limitations
- The official delivery time is explicitly premised on the required inputs and scope of work having been confirmed, and is limited to supported models and development boards; new architectures and new development boards typically take about one week.
- The official information does not provide specific pricing, charging model,free trial policy, or similar information.
- The official information does not provide a complete list of supported models, only the above examples of hardware platforms.
- Regarding data such as "about 5x improvement," the official information explicitly states that it comes from a specific demonstration (Isaac-0.2-2B + SAM2 on Jetson Thor) and does not represent general results for all models and hardware.
