Steel casting with AI — the porosity you find at machining gets paid for three times.

The problem

The defect gets into the casting at the pour — but it gets billed at machining or at the customer.

The quality of a casting is the intersection of four pieces of data that are almost never on the same screen:

1. What was poured — the furnace's chemical composition (carbon, manganese, silicon, sulfur, phosphorus). In the spectrometer system, usually with a delay.
2. How it was poured — pouring temperature, filling, mold parameters, atmosphere. In the molding plant's control system.
3. How it cooled — the thermal pattern at shakeout, the gradients. Almost nobody records it casting by casting.
4. What the casting looks like when it comes out — surface blowholes, veining, cold shuts, flash. The operator's visual inspection before shot blasting.

The quality manager knows it, but most defects are only confirmed at machining (when the internal porosity that wrecked a lathe operation shows up) or at the customer (when dye penetrant testing or ultrasound flags an inclusion). By then the cost is triple: the casting, the machining hours and the customer's trust. The classic system works 99% of the time. That other 1% is what the monthly committee pays for.

How it fits the IRIS system

iLEAN does not add a fifth system — it seals the cracks between melt, mold, shakeout and appearance.

The problem in a foundry is not a lack of information: it is information living on islands that, between the pour and the final inspection, never reaches the decision-maker cross-checked. iLEAN acts as the putty that fills those gaps, without asking you to change your spectrometer, your molding system or your classic inspection.

Edge looks at the casting with a thermal + visible camera at shakeout. Connect reads the melt and the mold wherever they live. The agent cross-checks the signature against the composition and the parameters, and flags the borderline casting before shot blasting. The person decides where it goes.

The three iLEAN pieces applied to steel casting:

• Edge — a terminal with machine vision (CNN) at shakeout. A thermal camera for a per-casting temperature map and a visible camera with raking light to classify surface defects. It flags the borderline casting and triggers an actuator (status light, diverting lane). It works with no network.
• Connect — captures the spectrometer composition, the mold parameters and the melt recipe, whether they come from a modern system, from an old machine or from a spreadsheet kept by the metallurgical lab. And it captures what arrives from outside: a change in the supplier's scrap, a new customer specification tightening S and P.
• Agent — cross-checks the thermal and visual pattern against the composition and the mold parameters. It recommends where each borderline casting should go (radiography, testing, controlled scrap) to the quality manager through whatever channel they use. The person decides and signs; no casting is discarded on its own.

See the full IRIS architecture →

Before and after

Classic inspection vs. cross-checked control with iLEAN

Aspect	Classic inspection + standalone spectrometer	With iLEAN Edge + Connect + Agent
Melt composition	In the spectrometer, almost never cross-checked	Attached to the specific casting at second zero
Thermal pattern at shakeout	Almost nobody records it casting by casting	Thermal camera per casting, cross-checked with composition and mold
Surface defects	Operator's visual inspection before shot blasting	Vision classified under raking light, before shot blasting
Porosity detection	At machining or already at the customer	Casting flagged as borderline for radiography or testing
Operation with no network	n/a	Edge keeps classifying on the panel's own light
File for the customer	Rebuilt by hand when a claim comes in	Per-casting dossier, automatic, with composition + thermography + photo

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.

• Carbon / stainless / alloy steel foundry, castings up to several hundred kg, supplying the mechanical, energy or rail sectors.
• Edge pilot at shakeout (thermal camera + visible camera + integration with the spectrometer and the mold system). First value expected in a few weeks.
• Reduction in castings whose defect is found at machining/at the customer, against your history: ≥ 30% — a conservative estimate.
• Indicative payback between 4 and 9 months, depending on the average cost per casting with machining hours wasted and the frequency of claims over the past year.
• The hard lever is every casting scrapped before machining (instead of finding it at the customer).

And the quality manager's reasonable doubt

“What if the AI flags a good casting as borderline?” — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI merely cross-checks thermal + visual signature + composition, the best models brought error below 1.5% ^[1]. And even so, the flagged casting goes to a second inspection (radiography, dye penetrant), not to the scrap bin. The person decides; no casting is discarded on its own. The three safety rings exist precisely for this.

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