EN 10204 3.1 certificates read by AI and checked against the order and the grade limits

What the AI does with a 3.1 certificate

An AI system can read EN 10204 3.1 inspection certificates from many steel mills, pull out the order number, grade, heat number, chemical composition and mechanical values, and check them against your order and the limits of the grade. The reading is done by AI models; the checking is done by fixed rules that your quality team writes, so a certificate passes only when every value is inside a limit you entered. On a small server with two NVIDIA RTX PRO 4500 Blackwell Server Edition cards, a day of incoming certificates takes well under an hour of computing time by our estimate.

This page is for the head of quality, incoming inspection or purchasing at a company that buys steel, stainless steel or other metallic products: machine builders, steel construction firms, stockholders, makers of pressure parts and fasteners. It explains the task in plain words first, then what the server needs, then the notes for whoever runs your IT.

What a 3.1 certificate is and why checking it takes time

EN 10204:2004 defines four types of inspection documents for metallic products. A declaration of compliance with the order (type 2.1) and a test report (type 2.2) are issued by the manufacturer on the basis of non-specific inspection, the second with test results. An inspection certificate 3.1 is issued by the manufacturer, states that the products comply with the order and gives the results of specific inspection, meaning tests on the delivered products or on the batch they come from. It is validated by the manufacturer's inspection representative, who is independent of the manufacturing department. A 3.2 certificate is prepared by that representative together with the purchaser's representative or an inspector named by official regulations. That is how published summaries of the 2004 edition describe them; the standard itself is sold by the standards bodies.

A second standard sets the vocabulary. EN 10168:2004 lists the information that may appear on inspection documents for steel products and gives each item a code. The A codes cover the commercial side, such as A 02 type of inspection document and A 07 purchaser order number. The B codes describe the product, such as B 02 steel designation, B 07 identification of the product and B 13 actual mass. The C codes hold the inspection results, with sections for tensile, hardness and notched-bar impact tests and for chemical composition, the D codes other tests, and the Z codes the validation.

The work in incoming inspection is the comparison. Someone has to find the order number, check that the grade and dimensions match the order item, read a table of ten or twenty element contents and compare each with the limits of the grade, check yield strength, tensile strength, elongation and impact energy at the required temperature, note the heat number so that the material can be traced later, and file the document. Every mill has its own layout, many write in two languages, and some certificates arrive as a scan of a fax or a phone photo.

Mill certificate extraction and the check against the order

The process has four steps, and only the first two use AI models.

1. Reading the page. An OCR model (optical character recognition, software that turns an image of text into text) converts each page into text and tables. Modern OCR models are small AI models trained on documents, and they keep the table structure, which matters for a composition table.
2. Filling a fixed form. A text model, the same kind of language model as a chat assistant, reads that text and fills a form with fixed fields: certificate type, order number, mill, grade, heat number, dimensions, mass, each element with its value, each mechanical value with its unit and test temperature. The form is a JSON schema, a strict list of fields and formats, and the inference software can force the model's answer to follow it, as our guide to vision-language models for document AI describes.
3. Checking by rules. A short program compares the form with the order from your ERP system and with a limit table per grade and per customer requirement. It flags every value outside a limit, every missing value, a type 2.2 document where the order asked for 3.1, and a missing validation by the inspection representative.
4. Human decision. Certificates without any flag can be released automatically if your quality manual allows it; everything flagged goes to an inspector with the page image and the value marked. The inspector releases, rejects or asks the mill.

A rule makes the decision because language models make occasional reading and reasoning errors, and a quality decision must be reproducible and auditable. A rule like "carbon at most the limit in our table for this grade" gives the same answer every time and can be shown to an auditor. The model's job is limited to reading. The limit table itself comes from the material standards and your customer specifications; whether the text of a standard may be loaded into an AI system at all is a licence question, which our guide to standards in company AI covers.

Reading errors are caught in three ways. The form keeps the page and position of each value, so the inspector sees the source in one click. Values that fail a plausibility rule, such as a sum of elements above 100 per cent or a tensile strength below the yield strength, are flagged even when they lie inside the limits. For critical parts, the page can be read twice by two different models, and any difference goes to a person.

Heat numbers and batch traceability

The heat number links a certificate to the melt the steel came from, and through it to the parts you make. Once the form is filled, the heat number goes into the batch record of your ERP or warehouse system together with the certificate number and a link to the archived PDF. When a customer later asks for the certificate of a part, or a complaint needs the material history, the search runs over structured fields instead of folders of scans. EN 10204 allows inspection documents to be kept and passed on electronically, as the German-language Wikipedia summary of the standard notes, so the archived PDF and the extracted record can replace the paper file if your quality manual says so.

The same pipeline reads the supplier certificates for other metallic products, castings and forgings as long as the document follows the same pattern. Certificates that come as data files rather than PDFs skip the reading step and go straight into the rules.

We supply the RTX PRO 4500 Server Edition and build servers with two or four of them for this kind of document work; send us a few typical certificates and your daily volume, and our team replies with a configuration.

The models and why they are small

This task does not need a large model. The text on a certificate is short, the fields are fixed, and the hard part, the comparison, is done by rules. Two open models show the scale.

Hugging Face model cards and "Files" pages of zai-org/GLM-OCR and Qwen/Qwen3-14B-FP8, read 25 October 2026.

GLM-OCR's card describes text, formula and table recognition and an information-extraction mode in which the prompt gives a JSON schema and "the output must strictly adhere to the defined JSON schema". It reports 1.86 pages a second on PDF documents at a single concurrency, without naming the GPU, and tags eight languages without listing them, so test it on certificates in each language your mills use. Qwen3-14B's card gives a native context of 32,768 tokens, far more than a certificate needs, and support for more than 100 languages and dialects. It thinks before it answers by default; for form filling, thinking is switched off with the documented enable_thinking setting. Other OCR models, such as DeepSeek-OCR-2 or olmOCR-2, are compared in our document AI guide, and the choice should follow a test on your own certificates.

Hardware: what the server needs

The certificates of a day are a small batch job. As an example, take 150 certificates a day with two pages each, 300 pages.

Our estimates. Users-per-card rule: 90 per cent of the 31.9 GiB the driver reports, less 3 GiB, less the weights. Form time from StorageReview's review of the RTX PRO 4500 Blackwell Server Edition (14 October 2026, vLLM 0.27.1): Ministral 3 14B FP8 with 8K input and 1K output flattens at about 180 output tokens per second from eight streams (its GPU overview page, modified 20 October 2026, lists 190; we plan with the lower figure); certificate volume and 800 output tokens per form are example values.

The OCR time uses a planning value of one page every five seconds, which we chose as a cautious figure. It is not a measurement: no one has published a page rate for GLM-OCR on this card. Scaled from Datalab's published rate for its larger 5B OCR model on an H100 by memory bandwidth, the card reaches at most about 0.34 pages a second, an upper bound by our estimate, and GLM-OCR has about a sixth of those parameters. A pilot with your own certificates gives the real figure.

The result is the point of the table: the daily certificates use the server for well under an hour. Suppliers' certificates rarely justify a server on their own. They justify one as part of the company AI server, which during the rest of the day reads order emails and PDFs into the ERP, drafts safety data sheets or supports quality management and audits.

Three ways to run it

Power by our estimate: cards at 165 W, two AMD EPYC 9135 for two cards or two EPYC 9355 for four, plus 15 per cent for the rest of the server, as in our guide to four RTX PRO 4500 Server Edition cards in 2U.

Inside an existing AI server. If your company already runs a server with RTX PRO 6000 Blackwell Server Edition cards, one card can be split by MIG (Multi-Instance GPU, a hardware partition of one card into isolated parts) into instances of 24 or 48 GB, as our page on RTX PRO 6000 MIG profiles explains. A 48 GB instance holds both models with room for many forms; in a 24 GB instance the 15.2 GiB text model leaves room for only a few forms by our rule, so the OCR model then takes a second instance.

Two RTX PRO 4500 Server Edition. NVIDIA's product page lists 32 GB of GDDR7 at 800 GB/s, 165 W and a passive single-slot design for this card. Two of them with two AMD EPYC processors and at least 128 GB of DDR5 ECC memory, twice the GPU memory as NVIDIA's guide for certified systems asks, make a first server for this and other office tasks. At about 0.84 kW by our estimate it fits one ordinary 230 V 16 A feed, which carries 2.94 kW at 80 per cent, but it belongs in a server room or a rack, not under a desk.

Four RTX PRO 4500 Server Edition. The four-card build with two EPYC 9355 and 384 GB, at about 1.4 kW by our estimate, still fits one 16 A feed, as our guide to GPU server power on 16 A, 32 A and three-phase feeds works through. It runs the document pipeline beside an internal assistant for several departments. A measurement on the delivered system confirms the power figure.

If you have no IT department, our guide to an AI server without an IT department describes who does what. We set up the models, the rules and the ERP connection through our Private AI/ML service; describe your certificates, your ERP and who checks them today, and we propose a pilot.

Notes for the IT team

1. Input. Certificates arrive by email, supplier portal or scanner. A mailbox rule or a watched folder feeds the queue; PDFs with a text layer can skip OCR, scans and photos cannot.
2. Serving. vLLM serves both models with an OpenAI-compatible API. Its documentation describes --gpu-memory-utilization as a per-instance limit and gives the example of two instances on one GPU at 0.5 each, so a single card can hold both engines; two cards keep them apart and are easier to run.
3. Form and rules. Keep the JSON schema, the limit tables and the rules in version control, with the quality manager as owner of the limit tables. Store each extracted value with page number and position, the model versions and the rule version, so that a decision can be reproduced later.
4. ERP link. The rules read the order item, grade and agreed requirements from the ERP and write the heat number, certificate number and result back to the batch record. Your ERP provider sets up that interface; we connect the AI side to it.
5. Access. Keep the engines on an internal network behind a gateway with keys, as our guide to an LLM gateway with keys and a query log explains.
6. Software on delivery. The server arrives on request with the operating system, the NVIDIA driver, CUDA and a container runtime. The RTX PRO 4500 Server Edition needs a driver of the R595 branch or later, as our RTX PRO 4500 Server Edition article sets out.

What we supply

You can buy this system from us: our team sizes the configuration to your workload and budget and replies within one business day. We build the servers to order with RTX PRO 4500 Server Edition or RTX PRO 6000 Server Edition cards, assembled and tested before delivery, delivered anywhere in the EU on one EU contract and invoice, with manufacturer warranty on every component handled through us; quantities and lead time are confirmed in writing with the quote. Our Private AI/ML service, with our engineering partner Vixen.UNO, deploys the OCR and text models, the checking rules and the connection to your ERP. Our guide to buying a four-card RTX PRO 4500 Server Edition server lists the build.

FAQ