Expertise is the bottleneck.
We ship it as a device.
IDSCOUT brings AI domain experts into Auto-ID practice. The first solution: an NFC encoding station that takes a job in plain language, backs every extracted value with the original quote, and only encodes after label 1 has been physically read back and confirmed in plain text. Customer data never leaves the device.
Pre-registered acceptance runs of the language layer, as of 16 Aug 2026. Thresholds fixed in advance, never moved. Live test under real conditions: 4/4 chip payloads byte-identical, triple-verified (station read-back, target comparison, independent smartphone dump).
Encoding NFC is easy. Setting it up is not.
Setting up RFID or NFC production today means knowing memory banks and lock modes, choosing EPC schemes with partition and filter values, managing access and kill passwords, and mastering NDEF record structure. Label software mirrors exactly that: dialogs built for specialists. For a label converter whose team comes from print production, this is the mountain NFC jobs fail on.
IDSCOUT turns the workflow around: the operator describes the job the way it reads on the job ticket. Vetted recipes and deterministic code do the rest.
This is what job setup looks like in typical label software today.
From job ticket to verified label
SimulationThree real job types, re-enacted. The demo shows the station's workflow: recognise the pattern, fill the form with an evidence traffic light, verify the first piece, only then run the batch. Green means literally evidenced, orange is a suggestion, red is missing. On the station itself, job analysis takes 8–13 seconds for deterministically recognisable jobs, 65–105 seconds with AI extraction — entirely on the device, measured under eval conditions.
First-piece check: label 1 has been physically read back
Job discarded. No further label was encoded. That is exactly what the gate is for.
UID↔Serial-Report: 200 rows · Audit-Log closed
Recipe/software version, model digest and spec hash logged per label
Two lanes. One rule.
“The AI never writes” is not a marketing line, it is the system's design rule. And it answers the question every specialist asks next: what if the model is wrong? Then the job dies at validation, guardrails or the first-piece check; the label never dies by the model.
Understands the job
- recognises the job pattern
- extracts values from the text, each backed by a literal quote
- answers domain questions extractively from local knowledge cards, with source references
Assembles and writes the job
- chip database and vetted recipes
- NDEF/raw builder, writing with full verify per label
- first-piece check as a mandatory gate before any batch
- audit log and UID↔serial report, offline
- 142 automated checks per release
One scout per domain
IDSCOUT is not one device but the AI layer above the IDCRAFT portfolio. Each expert solves a bottleneck where domain knowledge is missing today, with the same architecture: evidenced answers instead of claims, local instead of cloud, and an AI that structurally cannot break anything.
IDSCOUT Encode
The NFC encoding station. Plain-language job intake, evidence traffic light, first-piece check, full verify, audit report. For label converters who want to offer encoding as a service.
IDSCOUT Verify
The validation station: check instead of write. Goods-in, sampling, complaints, same evidence and reporting principle.
IDSCOUT Integrate
The integration assistant for developers: SDKs, protocols, sample code and debugging for the readers IDCRAFT distributes, grounded in vendor documentation, not hallucinated.
IDSCOUT Stage
The AI assistant for field engineers: commission and configure devices and run read tests. Measure transponders in the live application, measure field strength and tune the readers accordingly.
IDSCOUT Select
Selection guidance: which reader, chip and antenna for which use case, with traceable reasoning.
IDSCOUT Comply
Standards and compliance: EN 18220, ETSI EN 302 208, GS1/EPC, Digital Product Passport, answers with source references.
Where this leads
Long term, IDSCOUT grows into a standalone AI with multiple domain experts for the Auto-ID field. It is “consulting over box moving”, productised.
Local is not a feature. It is the precondition.
IDSCOUT Encode answers these questions by construction: the language model (Apache-2.0-licensed) runs on a Raspberry Pi 5 inside the device, the station is air-gap capable, CSV contents never reach the model, and the audit trail stays in-house. A device without a cloud connection answers most NIS2 supplier security questionnaires by itself. Based on our assessment the station does not fall under the high-risk use cases of Annex III of the EU AI Act; it visibly meets the Art. 50(1) transparency obligation in the operator UI. We provide our AI Act assessment to prospects as a one-pager. And the European alternative is growing: projects like Soofi are building open foundation models, developed and trained in Germany. IDSCOUT is itself a German development built on open weights: models stay interchangeable, including European ones, as soon as they fit the job.
Straight answers
Which chips and payloads does version 1 support?
NTAG213 plus ICODE SLI, SLIX and SLIX2. Payloads: NDEF link and text, digital business card (vCard), Wi-Fi access, Smart Poster, phone/e-mail/geo, raw blocks and customer-specific house formats — series fixed, counting or per label from CSV. XLSX import and NTAG password/lock features are on the roadmap.
Does the station need an internet connection?
No. The language model runs locally and the station works air-gapped. Remote maintenance is optional and remains a deliberate operator decision.
What happens when the AI is wrong?
It never writes, so an error cannot reach the label unchecked: a wrong suggestion fails at validation, at the guardrails, or at the latest at the first-piece check, where label 1 is physically read back. In our test runs no incorrect AI value ever reached a label.
Isn't this just read-after-write?
Read-after-write is good practice and documented by several printer manufacturers; we treat it as table stakes. The difference sits before and after: first-piece approval ahead of the batch and a UID↔serial report plus audit log fed back into the job, fully offline. Based on our market research of August 2026 we found no comparable product for this combination.
Who is IDSCOUT Encode for?
Version 1 targets label converters and printers in the DACH region who want to offer NFC encoding as a service without hiring RFID specialists. Operator documentation in German and English; the operator UI is currently German, English to follow.
Which language model runs on the station?
A locally operated model with open weights. The architecture is model-agnostic: models stay interchangeable, including European open-source models such as those currently emerging from the Soofi project, as soon as they fit the job. What matters is not which model runs, but that it never writes.
What does the station cost?
Price on request. Contact us for pilot terms.
Let us show you on your own job.
Bring a real job ticket. If the station does not understand it, we both learn something.
Request a pilot or demo