4,500 industrial cameras inspecting fabric, in real time, across 500+ machines in 7 countries. A rugged vision-and-AI stack, edge-first and field-proven from −10°C to +50°C. Starting in textile — expanding into every substrate where defects, waste and inefficiency cost the planet.
For: quality directors · plant operations leaders · manufacturing technology owners · manufacturers evaluating AI inspection platforms. Founder reads every demo request.
Surface inspection is a category, not a single product. The platform is built to handle any continuous or semi-continuous substrate — what changes between materials is the defect taxonomy, lighting setup and camera placement, not the underlying inference architecture. Below is what runs in production today, and what the architecture is being extended to next.
Six dedicated textile surface inspection products. Production scale: 4,500 cameras across 500+ machines, 7 countries.
The same edge inference, defect-detection and MES-integration architecture extends to other continuous or semi-continuous substrates. These are at varying stages of pilot, scoping or active research with prospective customers.
If your substrate is not listed, it probably extends. The platform is built to be re-trained on a new material class in weeks, not quarters. Email Harsha.
A non-exhaustive view of the defect classes the platform handles in production today, drawn from across the textile portfolio. Each one represents thousands of training examples across multiple customer sites. The taxonomy expands per substrate — defects on glass do not look like defects on knit fabric, but the inference architecture handling them is the same.
The most common knit defect. Detected at 30 cm of fabric, sub-50ms latency, on dyed and undyed material.
Subtle vertical defect from a degraded needle. Predictive signal — rising-frequency cluster indicates which needle will fail.
The hardest defect class. Core-spun yarn defeats IR sensors and cloud-based vision. The category Knit-I most strongly outperforms in.
Stains and foreign matter from machine maintenance or environment. Surface vs penetrated discrimination supported.
Yarn ruptures, broken courses, machine stops. Triggers auto-stop on critical events.
GSM drift over time, course density anomalies. Correlated with operator entry to surface MES gaps.
Multi-colour print mis-alignment, color drift over the print run, repeat pattern errors.
Shade variation across rolls, batch-to-batch consistency, dye uptake anomalies.
Cut-piece quality, edge uniformity, garment-stage defect detection.
Yarn cone density, package shape, surface contamination on the cone itself.
Mill-specific defects encountered across 4,500 cameras — each customer site adds to the shared defect knowledge base.
The platform is trained per site. New defect classes specific to your operation can be added during the pilot tuning window.
Surface inspection is not a model problem in 2026. The models exist. What separates a platform that scales to 4,500 cameras across 7 countries from one that gets stuck at the first site is the architecture around the model. Four decisions made early at CountAI continue to compound today.
All inference runs on an on-premise edge compute box at the machine. Sub-50ms latency, no cloud round-trip, no bandwidth ceiling, no data residency conversation with the customer's privacy team.
The system mounts at the take-down roll. It does not modify the machine. It ingests existing IP cameras via standard RTSP. The customer's MES, ERP and VMS continue to run untouched.
The same inference pipeline runs across Intel Core Ultra, NVIDIA Jetson, Hailo-8 / 10H and Raspberry Pi CM5 + accelerator. The chip choice can follow the use case — not the other way around.
Touchscreen at the machine, in the operator's local language. No data scientist required to interpret alerts. The hardest part of factory AI is not detection — it is the human side of deployment.
Start with a 30-minute call. I’ll walk you through real production examples from across the textile portfolio — faces blurred, customer identities anonymised — and we’ll map what surface inspection would look like on your line, your substrate, your cameras. No commitment afterwards. If it fits, we move to a paid 4-week pilot. If it doesn’t, you walk away with a sharper framework for evaluating any vendor in this category.
Founder reads every demo request. Usually replies the same day.
Computer vision running at the point of production to detect defects on a continuous or semi-continuous substrate — fabric, paper, metal sheet, glass, leather, food packaging — in real time. Unlike traditional end-of-line manual inspection or older optical inspection systems, modern AI surface inspection runs on edge hardware at the machine, detects defects as they form rather than after the roll is complete, and integrates with the customer’s existing MES and ERP without rip-and-replace.
Four things. First, scale — 4,500 cameras and 500+ machines in production across 7 countries. Second, substrate breadth within one vertical — six dedicated textile products covering knits, wovens, prints, cut piece, yarn cones and technical textiles. Third, chip-agnostic architecture — the same inference pipeline runs across Intel, NVIDIA Jetson, Hailo and Raspberry Pi compute, so deployments are not locked to one silicon family. Fourth, retrofit-first — ingests existing factory cameras, runs inference on-premise, no cloud dependency.
Production deployments today are in textile: circular knit fabric, woven fabric, printed and dyed fabric, cut-piece fabric, yarn cones, and technical textiles. The architecture extends to any continuous or semi-continuous substrate, and we are actively scaling into paper and pulp, food packaging films, glass and ceramic tile, leather, metal coil, and pharma packaging — each at different stages of pilot or research conversation.
No. The platform ingests existing IP cameras via standard RTSP / ONVIF streams. The only new piece of hardware is an on-premise edge compute box per machine or per site. Existing VMS and surveillance infrastructure continues to run in parallel — the AI layer is additive, not a replacement.
For the first batch of machines: typically under 24 hours of installation work, with operators trained on the local-language touchscreen on day one. Useful detection alerts start the same day. Tuning the per-machine model for the customer’s specific fabric mix, lighting conditions and defect priorities takes the first 30-60 days — that is real engineering work, not a setup wizard.
Traditional AOI systems excel in discrete inspection (PCB assembly, packaging endpoints) and are typically closed-stack solutions with proprietary CV. CountAI is designed for continuous-substrate surface inspection: the camera looks at material moving past it at speed, the model detects defects as the material forms or moves through the machine, and the platform is open at the integration layer (RTSP in, REST / OPC UA / MQTT out) rather than a closed black box.
Accuracy varies by substrate, defect class, camera placement and lighting — too much for a single global number to mean anything. What we commit to: a per-class precision and recall benchmark on your cameras during the pilot, measured against ground truth your team labels. If the numbers don’t justify the deployment, we don’t move forward.
Founder-written essays on the architecture, the chip landscape, the deployment economics, and how to evaluate vendors in this category.