“It Works Without Training”. Superb AI’s Push to Automate the Security Feed.
30 Oct 2026 · Stories from AIMX by Industrial Gemba’s Editorial Team Superb AI’s Product Owner Mitchell Kim on watching sites through existing cameras, a zero-shot foundation model, and...
Superb AI's Product Owner Mitchell Kim on watching sites through existing cameras, a zero-shot foundation model, and fine-tuning on a handful of images. 30 Oct 2026 · Stories from AIMX by Industrial Gemba’s Editorial Team
Superb AI’s Product Owner Mitchell Kim on watching sites through existing cameras, a zero-shot foundation model, and fine-tuning on a handful of images.
“Usually, there’s a human behind the camera watching everything,” says Mitchell Kim, Product Owner at Superb AI.
The South Korean computer vision company wants to change that. Rather than requiring expensive hardware overhauls, Superb AI plugs directly into the closed-circuit feeds ports, warehouses, construction yards, and municipal operators already run, turning passive security cameras into automated monitors for safety hazards, operational bottlenecks, quality defects, and regulatory compliance.
A Model Built for the Field
The engine behind this shift is ZERO, Superb AI’s vision foundation model built specifically for heavy industry.
Unlike general-purpose models trained on web-scraped imagery, Mitchell notes that ZERO was developed exclusively on industrial operational data: roughly one billion purpose-built data points. Because the architecture is zero-shot, operators can bypass the traditional onboarding lag.
“It works without prior training,” Mitchell explains, “and it works on your industry-specific domains right away.”
Ten Images, Not Ten Thousand
Edge cases, however, are inevitable on the factory floor. When off-the-shelf recognition misses an obscure part or unusual protocol, Superb AI takes a few-shot approach.
According to Mitchell, tuning the system requires only 10 to 20 site-specific images rather than the thousands historically needed to re-train computer vision pipelines. As edge devices log field hours, accuracy compounds incrementally without demanding a dedicated data science team on-site.
The Gemba Takeaway
For operations and site managers evaluating claims of instant deployment:
- Benchmark Before Training: Run the foundation model zero-shot against your raw archive. Your low-light feeds, glare angles, bad weather, and graveyard shifts. Measure baseline false positives and misses before committing to custom tuning.
- Audit the Few-Shot Lift: When zero-shot misses a critical hazard or defect, test the claim directly: supply exactly 10 to 20 labeled frames and demand quantified precision and recall gains on that specific variable.
This interview is part of Industrial Gemba’s coverage from AIMX – the region’s AI event featuring real-world use cases, innovation showcases, and strategic insights. Responses have been edited for length and clarity. Watch the full interview on YouTube.

