A GAFTA-Accredited Lab Tested Our AI Against Its Own Technicians. Here's the Review.

Most AI case studies are written by the vendor. This one starts with a company whose entire business is being the neutral party.

Sekargas Hamilton, UAB is one of the leading independent inspection and quality control companies in the Baltic region. Based in Klaipėda and part of the international J.S. Hamilton Group, it provides sampling, laboratory testing, cargo inspection and certification for agricultural and other commodity sectors, accredited to GAFTA, FOSFA International and ISO/IEC 17025. When an exporter and a buyer disagree about a cargo, Sekargas Hamilton is often the party whose numbers settle it.

Our clients are exporters, importers, traders, processors and logistics companies, for whom a reliable and independent quality assessment is critical. We work every day in an environment where data accuracy, traceability and trust are essential.

- Nikolaj Stankevič, Head of Inspections, Sekargas Hamilton

That is a demanding place to test an AI system. So we did the test their way: same samples, same standard, their technicians, their laboratory. This post covers both halves of the story: the review from the inspection side, in Nikolaj’s own words, and the measured results of the head-to-head test, including the parts that did not go our way.

Why an inspection company started looking at automation

Nothing about the pressure Nikolaj describes is unique to the Baltics. It is the same pressure every quality function in the grain trade is under: more data expected, faster, with a shorter tolerance for error.

Cargo and quality control processes are becoming increasingly complex, and clients expect faster access to information and greater data transparency. A large share of these processes has historically relied on manual data collection and administration, so we saw an opportunity to increase efficiency, reduce the risk of errors and ensure even better data traceability throughout the inspection process.

- Nikolaj Stankevič, Head of Inspections, Sekargas Hamilton

The important word there is traceability, not speed. In an accredited inspection environment, a result that cannot be reconstructed later is worth very little. That framing shaped what we were asked to prove.

What they looked for in a partner

We were looking for a partner who understands not only technology but also the specifics of grain trading and quality control. It was important to us that the solution matched real inspection processes, ensured data integrity and traceability, and could be developed together with business needs.

- Nikolaj Stankevič, Head of Inspections, Sekargas Hamilton

This is the part vendors tend to underestimate. An inspection lab does not buy a camera. It buys a method it will have to defend to a client, an auditor, and occasionally an arbitration panel. Our job in the pilot was less about demonstrating a model and more about fitting inside a procedure that already works.

The test: 4 technicians, 10 samples, 21 categories

The pilot ended with a structured head-to-head comparison, designed by the laboratory rather than by us.

• 10 wheat samples, each analysed across 21 impurity categories
• 4 experienced Sekargas Hamilton technicians, plus GrainODM
• Roughly 840 individual measurements
• Method: EN 15587, in two modes
  • Full analysis, ~30 minutes per sample
  • Rushed analysis, ~8 minutes per sample, simulating peak-season conditions
  • GrainODM: ~3 seconds per sample
• Each technician performed an equal number of tests in both modes, so mode effects could be separated from individual skill
• One sample was given twice to the same technician, as a repeatability check
• The reference value for every measurement was the average of the four technicians

There is no independent ground truth in impurity analysis. There is only the human reference, which is why the test measured the technicians against each other as carefully as it measured the machine.

Result 1: the AI was the most consistent participant, by a clear margin

Every impurity fraction in a sample should sum to roughly 100%. How tightly a participant holds that line across ten different samples is a direct measure of internal stability.

Participant	Range of sample totals	Spread
GrainODM	98.87% – 100.01%	1.14pp
Technician 1	98.57% – 100.00%	1.43pp
Technician 4	98.20% – 100.36%	2.16pp
Technician 2	97.66% – 100.48%	2.82pp
Technician 3	95.60% – 99.78%	4.18pp

GrainODM was the tightest participant in the test. That matters operationally more than it sounds: it means the same consignment measured twice, on two different days, by two different shifts, produces the same number. There is no “difficult morning” factor in a model.

It is also worth being precise about what this metric is not. Consistency is not accuracy. A system can be reliably stable and still be reliably off. Which brings us to the next result.

Result 2: on accuracy, the average technician was closer to the reference

Measured as mean absolute error against the four-technician average:

Participant	MAE vs. reference	Measurements
Technician 2	0.180pp	209
Technician 1	0.223pp	210
Technician 4	0.229pp	210
Technician 3	0.266pp	210
GrainODM	0.491pp	208

We are publishing this table because leaving it out would make the rest of the post untrustworthy. On this test, on this grain flow, GrainODM sat further from the human mean than any of the four technicians did.

Two pieces of context belong next to that number, and neither of them cancels it.

First, the average spread between the four technicians on a single measurement was 0.66pp - larger than GrainODM’s deviation from their mean. The reference itself carries substantial noise.

Second, the aggregate figure hides very large differences between categories. Which is the actually useful finding.

Result 3: 7 of 21 categories are already at laboratory level

In a third of the impurity categories, GrainODM’s measurements are practically indistinguishable from the technicians’:

Impurity	Technician average	GrainODM	Difference
Ergot	0.001%	0.008%	0.01pp
Stones	0.007%	0.029%	0.02pp
Husks	0.061%	0.025%	0.04pp
Maize	0.415%	0.419%	0.05pp
Oats	0.227%	0.251%	0.05pp
Other impurities	0.124%	0.114%	0.08pp
Barley	0.989%	0.941%	0.08pp

These are not the easy categories. Ergot and stones are exactly the contaminants an inspection company cannot afford to miss, and the AI measures them inside the human noise band today.

Result 4: where it still gets things wrong

Six categories carry a bias large enough to matter commercially:

Impurity	Technician average	GrainODM	Bias
Sound grains	88.93%	86.22%	-2.70pp
Darkened grains	0.354%	2.620%	+2.27pp
Small / shrivelled kernels	1.588%	0.526%	-1.06pp
Damaged grains	0.059%	0.738%	+0.68pp
Fusarium	0.452%	0.983%	+0.53pp
Triticale	0.624%	1.140%	+0.52pp

The pattern is not random, and that is the encouraging part. Darkened, damaged and fusarium-affected kernels are one visual family - overlapping colour and texture cues that the classifier is currently separating too aggressively. Sound grains then absorb the residual as a mirror image of those errors.

There is a directional consequence worth stating plainly: the bias is one-sided against the supplier. Where GrainODM and the technicians disagreed on a quality class, the AI assigned the stricter class in the large majority of cases. A system that errs toward rejecting good cargo is not a neutral system, and it is not something an inspection company can deploy as a final verdict.

The fix is not a general model improvement. It is facility-specific training data: 50 to 100 real samples from the Sekargas Hamilton grain flow, weighted toward those categories, then targeted fine-tuning and a re-test on the same protocol with a statistically adequate sample count. That work runs 4 to 8 weeks.

Result 5: the reference has more noise than the industry admits

The most uncomfortable finding of the test was not about the AI.

• The four technicians agreed on the quality class in only 66% of cases. In a third of cases, at least one disagreed with the rest.
• On sound grain content, the spread between technicians on the same sample reached 5.69pp (94.16 / 88.60 / 91.20 / 88.47).
• On one sample, the fusarium class assigned by the four technicians was 1, 1, 4, 4: two experienced professionals called it best quality, two called it worst. GrainODM returned 1, agreeing with the first pair.
• A rushed 8-minute test was, in this sample, marginally more accurate than the full 30-minute one (MAE 0.188pp vs 0.222pp). At n=10 that may be a real focusing effect or an artefact, but it does not support the assumption that more time automatically means a better number.

None of this is a criticism of the laboratory. It is an honest picture of what visual impurity analysis is: a skilled human judgement with real variance, performed under time pressure, on a physically heterogeneous material. It is also the reason a stable, documented, repeatable measurement has value even when it is not yet the most accurate one in the room.

Where this leaves the deployment

Given the results above, the honest positioning is narrow and specific.

What GrainODM is used for today:

• Pre-screening at intake. A 3-second measurement on every load, triaging consignments into clear, borderline, and check-manually. Technician time goes to the samples that need judgement.
• Continuous monitoring. Every consignment gets a measurement and a photograph, rather than only the ones selected for full analysis.
• Audit trail. A timestamped, reopenable record with every object segmented and classified. This is documentation that manual grading simply did not produce.

What it is not used for:

• An autonomous replacement for laboratory analysis, particularly in the categories with a known bias. Those stay with the technician until calibration closes them and a re-test confirms it.

The review from the inspection side

Nikolaj’s assessment of the pilot process itself, on the record:

The pilot stage was constructive and results-oriented. The GrainODM team worked actively with our specialists, responded quickly to observations and sought to understand our real working processes. We value the fact that the solution was improved on the basis of practical user experience, and that communication throughout the project was open and effective.

- Nikolaj Stankevič, Head of Inspections, Sekargas Hamilton

On where the value lands for their own clients:

We see the greatest benefit for clients in data transparency, faster access to information and better traceability. Digitalised processes make it possible to share results more quickly, ensure consistent data management and provide more confidence in the decisions being made. In grain trading, where quality data often becomes an important part of commercial decisions, this is a significant advantage.

- Nikolaj Stankevič, Head of Inspections, Sekargas Hamilton

And his answer to the question we asked last, which is the one most quality managers are actually sitting with:

Process automation today is becoming not a competitive advantage but a natural direction of business evolution. The most important thing is to start from clearly identified processes and to choose a partner who understands the specifics of your operation. Our experience shows that properly implemented digitalisation increases efficiency, reduces the administrative burden, improves data quality and creates greater value for clients.

- Nikolaj Stankevič, Head of Inspections, Sekargas Hamilton

What we took away from it

A vendor-run demo would have shown the seven categories and stopped. A test run inside an accredited inspection laboratory showed all twenty-one, and the six that are wrong are more useful to us than the seven that are right.

Three things we would repeat at any facility considering this:

1. Test against multiple technicians, not one. A single reference hides the variance that determines what “accurate” can even mean at your site.
2. Report per category, not in aggregate. A single headline accuracy figure is close to meaningless when category sizes span four orders of magnitude.
3. Deploy to the use case the data supports. Pre-screening and traceability are defensible on this evidence today. Autonomous grading in biased categories is not, and claiming otherwise is how pilots turn into disputes.

The calibration work at Sekargas Hamilton continues on real samples from their grain flow, with a defined re-test at the end of it.

Related reading: our validation against five lab technicians on 600+ wheat tests, the Allive hemp inspection case study, and a primer on EN 15587 besatz analysis for wheat.

---

Quotes from Nikolaj Stankevič were provided in writing in Lithuanian and are published here in translation. Test data: Sekargas Hamilton final testing round, April 2026, 10 wheat samples across 21 impurity categories, EN 15587.