ZETICVision

Real-time vision AI on edge hardware

Detect sooner.
Miss nothing.

We make your vision model fast enough to see every frame on the hardware you deploy. No cloud link, no dropped frames, more time to react.

What ZETIC does

From your model to
measured on-device performance.

  1. Your model
    & hardware

  2. Automated
    optimization

  3. Tailored
    inference engine

Recorded on the hardware

Same board.
Before and after.

Point cloud, stereo depth and object detection, recorded on the real board.

Non-vision demos: VLA, LLM and more
VLM + segmentation / Jetson ThorRecorded run

Prompt-to-tracking response

10.5× faster at the peak frame

Baseline 3.16 seconds. ZETIC 0.30 seconds.

Measurement scope

The difference

Same model. Same chip.
Only how it runs changes.

Vendor toolchains convert your model and run it on general-purpose kernels. We design the execution for your model on your chip.

  • 01Model graphWhat must run
  • 02Native kernelsWritten for the hot path
  • 03Runtime + stateMemory, scheduling, launch order
  • 04Target siliconMeasured on the real board
01 MODEL GRAPH02 NATIVE KERNELS03 RUNTIME + STATE04 TARGET SILICON
Conceptual execution stack

Case study · In production

Vision AI for quality control
on a semiconductor line.

On-device vision AI on a live production line, running on embedded hardware.

CustomerLeading memory
semiconductor company
  1. InputLive wafer images

    Cameras over the wafer line.

  2. On-device AIVision detection model

    Finds defects on each wafer, on the device.

  3. OutputQuality decision

    Good wafers move on; defects are flagged.

Defects caught earlyFound in-line, before a bad wafer runs through every remaining step. That processing cost is saved.
What it tookLimited production data.

Case study · Robotics

Language-guided object tracking
in real time, on a robot.

A top electronics company's robotics team. A VLM finds the object from a spoken instruction, and SAM2 tracks it.

  • 10.5×Faster lock-on3.16 s → 0.30 s at the peak frame
  • 30+ fpsReal-time trackingEven while the VLM answers
  • 2×+Faster than a YC startupA Y Combinator distillation startup, per the customer's own validation
InputVideo + instruction On-device · Jetson ThorVLM, then SAM2 OutputCoordinates + tracking Delivered as a ROS2 package

What we deliver

An optimized package.
Measured results.

  • Optimized package

    Ready to run on your board.

  • Integration artifacts

    To connect it to your software.

  • Performance report

    Latency, memory and power, reference vs optimized.

How an engagement runs

  1. 01Scope

    Model, target

  2. 02Profile

    Baseline

  3. 03Optimize

    On the board

  4. 04Validate

    Result, handoff

Trusted by

Teams shipping vision on devices.

Customers

Samsung Research America LG Electronics SK hynix

Technology ecosystem

Qualcomm Advantech Infineon Liquid AI

Programs and community

NVIDIA Inception EDGE AI Foundation TechCrunch Startup Battlefield 200

About ZETIC

Hardware-aware
inference optimization.

We help teams run their AI models on the hardware they ship. Founded in April 2024.

Team experience across research and industry

Qualcomm Amazon Deloitte Stanford University EPFL Seoul National University KAIST Northwestern University

Contact: seongjun@zetic.ai