What we do
We focus on two areas where Vision-Language Models create real value on the floor — and we deliver each one end to end, from optics and cameras to AI and integration.
Read every label, integrate every result
A combined build of liquid lens × line camera × VLM × WMS integration. The VLM reads labels, waybills and lot numbers — across different fonts, layouts, fine print and handwriting — with no master registration, and matches them against your WMS master data. A movable-part-free liquid lens keeps parcels of varying height in focus, so the line never stops. Inference runs on the edge, on-premises.
Hybrid inspection that fits high-mix lines
A rule-based × CNN × VLM hybrid. CNN and rule-based processing handle real-time inference at line speed, while the VLM works behind the scenes — generating NG (defect) images, auto-annotating, and supporting training — to solve the chronic "not enough defect samples" problem. A browser-based training UI lets floor operators run it, and we can add a VLM layer on top of your existing image-processing system.
Why Nsight
VLM cuts training cost
VLM-driven NG-image generation, auto-annotation and a browser-based training UI make high-mix inspection — once uneconomical to automate — practical.
Ex-Keyence optics know-how
We design lighting, camera placement and the inspection flow together — hardware, software and inspection know-how as one.
Start small, decide with data
A free sample evaluation takes days; a PoC can start in as little as two weeks. Continue or scale based on numbers, not promises.
Our hybrid architecture: rule-based × CNN × VLM
The hard part of multi-SKU and AI visual inspection is the design decision — which technique to use, where, and how. Nsight combines three layers and switches the VLM's role by task.
Inspection tasks (high-mix, variable height, etc.): the VLM is kept off the real-time loop for latency reasons and instead generates training data, annotates new SKUs, and assists on ambiguous cases. CNN × rule-based handles production inference at line speed.
OCR / label-reading tasks (logistics labels, engraving matching, etc.): the VLM runs direct inference, understanding text position and meaning with no training and matching it against master data — robust to layout changes, multiple languages and handwriting.
Optical know-how from Keyence's image-processing division — lighting, cameras, lenses and inspection flow designed as one — produces the clean image input that makes AI easier to train.
How we work
Site visit & interview
We see the line in person — SKU count, flow, lighting, existing equipment — and pick candidate camera positions together. (~1 day)
Plan & free evaluation
We run a free evaluation and produce a PoC design covering inspection method, KPIs and design. (1–2 weeks)
PoC
We validate on real equipment, accumulate logs and footage, and review results monthly. (1–3 months)
Rollout & expansion
Proven patterns move to production and roll out to other lines and sites. (scalable)