Phenolic plywood formwork inspection — a defect that transfers to exposed concrete is an urgent replacement on site.

The problem

The site customer does not see the statistics — they see the specific panel, from 50 meters away.

A phenolic plywood formwork panel lives four lives: you press it, you ship it, it travels to the site, it is assembled as formwork and, once the forms come off, its imprint stays in the concrete. If everything goes well, it lasts dozens of cycles. If a defect slipped past manual sampling, it shows up when the wall is standing and the civil works customer (general contractor, site engineer, owner) points at it from 50 meters away: there is an adhesive mark here, a raised blemish there, and over there you can see the wrinkles in the phenolic film.

Classic inspection works statistically — targeted sampling, the operator's eye, the quality manager's judgment — and that is usually enough for 99% of the panels. The problem is that with exposed concrete the remaining 1% is not an incident: it is an urgent replacement on site, with a compound cost:

1. The product received on site that gets rejected at stripping, with reverse transport or destruction.
2. The formwork assembled with that panel, concrete poured, defect invisible until the mold is removed.
3. The urgent replacement with stock product or prioritized production, to keep the site from stopping.
4. The damage to the account — the general contractor who asks for different terms on the next contract, or switches supplier.

Sampling is not the problem; the problem is that sampling lets through whatever goes onto the pallet between samples, and that pallet is hundreds of kilometers away before anyone finds out. Inspecting 100% in line changes the economics of shipping.

How it fits the IRIS system

iLEAN Vision does not replace the quality manager — it seals the cracks between his judgment and each specific panel.

The problem is not a lack of judgment: your quality manager knows perfectly well which defects to reject. The problem is that one person cannot look at every panel, panel by panel, face A and face B, across three shifts. iLEAN acts as the putty between quality's criteria and the line, without asking you to change the press or the sander.

Vision sees every panel with your team's criteria. Connect reads the batch's destination (exposed concrete formwork vs. standard formwork). The agent classifies, holds the critical ones and leaves the dossier behind. The person signs — never the other way round.

The iLEAN pieces applied to phenolic plywood inspection:

• iLEAN Vision (Edge) — a terminal with machine vision (CNN) and industrial cameras with grazing light (to capture relief in the phenolic surface) and reflected light (to capture adhesion and adhesive marks). It classifies panel by panel, face A and face B, labels each defect critical/cosmetic/acceptable according to your team's real criteria, and triggers the diverting actuator. It works without a network: as long as the terminal has power, the CNN runs inference locally.
• Connect — reads the batch's destination in the MES or ERP (exposed concrete formwork vs. standard formwork) and applies the right acceptance threshold. It also captures what arrives from outside the system (a site customer's specification by email, a change of criteria for fine-finish concrete) at second zero.
• Agent — cross-references the panel image, the batch destination, the customer's criteria and the line's history. It builds the per-pallet dossier with a photo of every panel and the reason for its classification; if the line builds up an anomalous pattern (more bubbles than usual, a lighting drift), it alerts the shift lead on whatever channel he uses. The person validates the borderline cases and signs; the pallet does not leave on its own.

See the full IRIS architecture →

Before and after

Manual sampling + operator judgment vs. 100% inspection with iLEAN Vision

Aspect	Manual sampling	With iLEAN Vision + Connect + Agent
Inspection coverage	Targeted sampling, borderline cases by eye	100% of the panels, face A and face B
Critical vs. cosmetic defect	Judgment of the operator on shift	CNN trained on the quality team's criteria
Adapting to the batch destination	Written procedure, depends on who passes the word	Connect reads the destination and applies the threshold
Detection before shipping	At the final pallet or on return from site	In line, panel by panel
Operation without a network	n/a	Vision keeps classifying as long as the cabinet has power
Evidence for the contractor	Manual reconstruction of the pallet	Photo per panel + classification + reason

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific data of your plant. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Phenolic plywood plant with one press+sanding line, a mix of standard formwork and formwork for exposed concrete, with documented returns from sites over recent years.
• Vision pilot at the press or sander outfeed (industrial camera with dedicated lighting + Edge terminal with CNN + sorting actuator + integration with the MES). First value expected within a few weeks, with the first run classified automatically.
• Indicative payback between 4 and 9 months, depending on volume, the historical frequency of on-site replacements and the share of production destined for exposed concrete.
• The hard lever is every urgent replacement avoided: product pulled from the site, reverse transport, schedule stoppage, risk to the account. Expected reduction of returns for visible-face defects ≥ 30% against the baseline.

And the quality manager's reasonable doubt

“What if the AI approves a panel with a critical defect?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely compares an image against closed criteria learned from your team, the best models brought the error below 1.5% ^[1]. And even so, what is critical is never decided alone: iLEAN holds the borderline cases and the person signs. The three safety rings exist precisely for this.

^[1] OpenAI paper “Why Language Models Hallucinate”, 2025 — on the reliability of AI in anchored tasks.