DSTEK KOREA

Physical AI engines, built in Korea

A large model for the physical world — from industrial vision to embodied intelligence.

See the technology Get in touch
  • Seoul, Republic of Korea
  • Incorporated 2025
  • Founded by SUALAB and Sendbird alumni

THE PROBLEM

Machines still cannot see

The bottleneck in manufacturing is perception, not automation

  1. Rule-based vision breaks

    Traditional machine vision needs a hand-tuned recipe per product, per line. Every change means re-engineering.

  2. Models do not transfer

    A model trained on one part fails on the next. Each deployment restarts data collection from zero.

  3. Robots lack understanding

    Automation can move, but cannot judge. Without perception, physical systems stay blind and scripted.

WHAT WE BUILD

From inspection to physical intelligence

We started in vision software. We are now building the model layer beneath it.

NOW

Vision inspection engine

Automated defect detection, classification and dimensional measurement running on live production lines.

BUILDING

Physical AI large model

A general perception model for the physical world — trained once, adapted to new parts and processes with minimal data.

NEXT

Robotics & embodied systems

Perception paired with actuation: machines that judge and act, not just execute scripted motion.

The same perception core powers all three. Inspection funds it; physical AI scales it.

THE MODEL

Our large model for physical AI

Trained on real production data, not benchmarks

What it does

  • Generalizes across parts, materials and lighting
  • Learns a new inspection task from few samples
  • Outputs judgment, not just pixels — pass/fail, defect type, location
  • Runs at production cycle time on the factory floor

Why we can build it

  • Production-proven vision stack already deployed in factories
  • Direct access to real defect data through customer co-development
  • Founding team with deep-learning vision background from SUALAB
  • Korean manufacturing base as the training and validation ground

Data advantage: we develop on-site with our customers, so the model trains on the exact conditions it must survive.

MARKET & TRACTION

Korean manufacturing as the beachhead

Semiconductor, PCB, automotive and battery — where inspection is hardest and most valuable

SECTORS WE SERVE

  • Semiconductor
  • PCB
  • Automotive
  • Battery

Engagements include on-site software development with the customer.

2025 Incorporated in Seoul
Founding team from SUALAB and Sendbird
Enterprise engagements in progress

Co-development on the factory floor is our moat — it produces data no dataset can buy.

TEAM

The team

Founding members from two of Korea's defining AI and software companies

CEO & CTO

Gyuyoung Hwang

Leads product and engineering direction. Early member at SUALAB and Sendbird.

Head of Engineering

Yong Kim

Owns the vision and physical-AI model stack. Early member at SUALAB.

Head of Sales

Dongsung Lee

Runs on-site deployment and customer co-development in Korean factories.

SUALAB — Korea's leading deep-learning machine vision company. Sendbird — Korean-founded global SaaS platform.

  • Seoul, Korea
  • Incorporated December 2025
  • Engineering team actively expanding

COMPUTE

How we use NVIDIA

Today in production, and at far greater scale from Q4 2026

TODAY

NVIDIA GPUs for model training and for edge inference inside deployed inspection systems on customer lines.

IN PROCUREMENT

An 8-GPU NVIDIA B300 system to train and serve our own large model on-premises in Korea. Purchase decision in August 2026.

What Inception unlocks for us

Hardware
Faster, better-priced access to Blackwell-class compute at the exact moment we scale training.
Ecosystem
NVIDIA AI Enterprise, DLI training and technical guidance as we move from vision models to physical AI.

Compute is the binding constraint on our roadmap. Everything else is already in place.

ROADMAP

Roadmap

From deployed inspection to a general physical AI engine

  1. 2026 H2

    Deploy GPU infrastructure

    Install 8-GPU B300 system. Begin large-model training on production inspection data.

  2. 2027 H1

    Ship the model

    Roll the physical AI model into live customer lines. Expand engineering team.

  3. 2027 H2

    Enter robotics

    Extend perception into embodied systems for manufacturing automation.

CONTACT

Talk to us

Manufacturing inquiries, partnerships, and engineering roles.