Human inspectors miss 20–40% of fabric defects at production speed. Our TensorRT-optimised Vision Transformer pipeline detects them in milliseconds — on factory-edge NVIDIA hardware. Built for the 3.6M textile mills that humans inspect today.
At production speed — 30–120 metres of fabric per minute — even trained inspectors miss between one and four out of every ten defects. The cost compounds downstream as garment rejections, brand returns, and second-grade resale.
Cameras over the loom feed a TensorRT-compiled Vision Transformer running on a single NVIDIA edge GPU. Detections are flagged on-line with bounding boxes and defect class — no cloud round-trip, no production-line stop.
No retrofitting. From training on cuDNN-accelerated Vision Transformers to edge-deployed TensorRT inference under Triton, every layer is NVIDIA-native.
The textile value chain — fibre, yarn, fabric, garment, retail — flows through spinning mills, weaving units, dye houses, and garment factories. We start at the fabric stage, where defects cost the most and humans miss the most.
| Capability | FabricScan | AWS Rekognition | Google Vision |
|---|---|---|---|
| Textile-domain training data | ✓ | Generic objects | Generic objects |
| Sub-100ms factory-edge inference | ✓ TensorRT | Cloud round-trip | Cloud round-trip |
| No-cloud / on-prem deployment | ✓ NIM | Cloud only | Cloud only |
| SME-mill pricing | ✓ | Per-image enterprise | Per-image enterprise |
| Per-customer fine-tuning | ✓ NeMo | ✗ | Limited |
| Founder-led textile distribution | ✓ 10yr network | ✗ | ✗ |
We are taking on a limited number of textile-mill pilot partners for the Q2 2026 FabricScan TensorRT deployment. If you operate a textile mill anywhere in the world, or you sell to the textile sector — email the founder directly.