Arc vision, acoustic analysis and voice advisory in the cell. AI advises, the operator decides, the PLC controls.
Less steel, wire and gas wasted at the arc — and every weld becomes machine-readable material-passport data. Digital monitoring & analytics for metal fabrication · ESPR-ready by design
Market Context
The skilled welder shortage was already structural. Mobilization, retirement and EU reshoring pushed it past the breaking point. The knowledge that keeps rejects low is walking out of the shop, and the rejects are made of steel.
European Welding Federation projects a structural shortfall of 330,000 qualified welders by 2027 – driven by retirement, reshoring, and the energy transition. CEE manufacturing absorbs the gap first.
A single 5-day MIG/MAG course at Polski Instytut Spawalnictwa or SLV runs €1,500–2,500 plus travel and lost production. For a 12-person shop replacing two welders a year, that's €10K+ disappearing into off-site training that may not transfer to your specific robot.
Tier-2 fabricators in Poland, Romania and western Ukraine report 40% YoY growth in EU subcontract orders. The bottleneck is no longer demand – it's qualified operators standing next to a six-figure robot they can't yet program confidently.
Raw-Materials Impact
A rejected weld is not just a quality event — it is steel, filler wire, shielding gas and arc energy already spent, and spent again on the rework.
Rework consumes the base metal twice and the filler wire twice. Preventing the defect at the arc is the only point where the material is still recoverable.
Ar/CO₂ mixtures and arc energy scale linearly with arc-on time. Less rework means less arc-on time, and lower embodied carbon per part.
Scrapped weldments enter the recycling loop prematurely. ARCMENTOR keeps material in the use phase — the highest-value loop in the EU Circular Economy Action Plan hierarchy.
⚠ Indicative model, not measured pilot data. ARCMENTOR is pre-pilot. Figures are per robotic cell per year, for each single percentage point of avoided rejects. Assumptions, formulas and sensitivity are published in docs/business/material-savings-model.md.
modelVersion: v0.1.0-model
What It Is Worth
Saved material and avoided rework labour are computed by two separate models with two separate sets of assumptions. Both are published, both are pre-pilot, and neither is allowed to borrow the other's number for a headline.
Steel, filler wire, shielding gas and arc energy that a prevented reject never spends twice. This is the raw-materials case, and it is deliberately small.
Gouging, re-welding and re-inspecting a rejected weld. This is the case a CFO recognises — 3× to 17× the material figure, and the reason saved kilograms alone do not justify an industrial AI system.
The per-cell figure is modest; the aggregate is not. One hundred cells, one percentage point of avoided rejects, one year — steel kept in the use phase instead of going back to the melt.
⚠ Two independent models, deliberately kept apart: docs/business/material-savings-model.md (material) and docs/business/copq-model.md (rework labour). Both rest on the same unearned assumption — one percentage point of avoided rejects — which a pilot has to earn before either figure is quotable. Summing them into a single benefit number is forbidden by both documents.
modelVersion: v0.1.0-model · v0.1.0-copq
Regulatory Tailwind
The EU is phasing in the Ecodesign for Sustainable Products Regulation. Iron and steel is the first prioritised product group; aluminium follows. No product-specific delegated act is in force yet — so nobody can be compliant today, and anyone claiming to be is guessing.
Fabricators will be asked for product-level material and process data they do not have. Today it lives in ERP exports, spreadsheets, and the memory of one quality engineer.
| DPP field | Who owns it | ARCMENTOR today |
|---|---|---|
| Unique product identifier | customer ERP | — not ours |
| Material composition | supplier | — not ours (mill certificate, EN 10204 3.1) |
| Welding parameters | robot controller | — not ingested today |
| Acceptance level (EN ISO 5817) | certified IWE / IWS | — never ours: a certified human act |
| Part identity (weld → part) | customer ERP | blocking gap — a DPP is product-centric; we are cell-centric |
| Weld process observation | ARCMENTOR | ✓ arc vision, per weld |
| Measurement provenance | ARCMENTOR | ✓ sync quality + clock drift, per event |
| Substantiating audit trail | ARCMENTOR | ✓ append-only, metadata-only |
ARCMENTOR certifies nothing. It produces the process-evidence layer that is hardest to reconstruct after the fact — and it says plainly which fields it does not produce. The seventeen rows we cannot fill are the honest part of this table.
System Boundaries
Three levels, stitched together by one weld id. The boundary between them is not a diagram convention — it is what keeps an advisory system from quietly becoming a control system.
A PLC owns the authoritative arc-on signal, meters wire, gas and energy per weld, and drives the only indicator lamps in the cell. Its analogue outputs are left physically unwired, so no current or voltage setpoint can be written at all.
Camera and microphone go straight into the edge box, past the switches. Vision and audio are synchronised on one monotonic clock, fused by deterministic rules, and spoken into the operator's headset without a single network call.
Diagnostic answers grounded in welding standards, each one passed through a deterministic safety filter that fails closed: if the filter is down, the user gets an error rather than an unfiltered model response.
How it works
Lucid Triton IMX490 HDR camera on the torch + miniDSP UMIK-1 calibrated microphone. Two modalities, timestamped on a single edge clock and fused only while drift stays within ±20 ms — beyond that the system degrades to vision-only rather than guess.
2 modalities · single edge clock · ≤20 ms driftEdge AI inference (YOLOv8 on Jetson Orin) classifies arc state in real time: porosity risk, undercut, lack of fusion, torch angle, travel-speed deviation. Design target: p95 under 300 ms – faster than the operator's reflex. Not yet measured on target hardware.
edge inference · target p95 < 300 msVoice prompts in Ukrainian, Polish or English through a bone-conduction headset under the welding hood: "drop the angle 5° right" · "slow down 20%" · "stop – back-step and re-strike". Senior welder Human-in-the-Loop validates novel cases.
native UA/PL/EN · sub-second cueEvery successful weld auto-logs to a WPS record: parameters, joint geometry, operator, traceable serial. Cloud-side nightly retraining of the defect model. Audit-ready for ISO 3834 / EN 15085 acceptance.
WPS · ISO 3834 · EN 15085Solutions
Built for fabrication shops that already own a welding robot but can't find or keep experienced operators. Built for any MIG/MAG welding robot cell.
Deployed and running in the cloud today.
Ask in plain language and get welding-parameter suggestions and defect diagnosis from the cloud agent – with cited sources, and a senior-review flag when the evidence is thin.
Live · REST + WebSocket APIEvery AI answer passes a deterministic safety filter before it reaches the user: imperative equipment commands are masked and unsourced claims are flagged for senior review. The AI gives recommendations only – never equipment commands.
Live · always-onEU data residency by design – all cloud resources run in eu-central-1. Privacy-by-default is built in: no raw audio leaves the device and the audit trail is append-only.
Live · EU data residencyPlanned for Wave 2–3 (edge) – not yet available.
Will deliver real-time voice coaching during training shifts – a co-pilot that watches the arc, hears the bead and prompts the operator to correct within seconds. Advisory only: the operator stays in control.
Planned · voice coaching (Wave 2)Will turn every weld into an auditable record – parameters, joint, plate, operator, time-stamp – and generate the WPS book an EU quality auditor expects, without a quality engineer spending days per part.
Planned · WPS automation (Wave 2)Will flag likely porosity and lack-of-fusion from acoustic and visual signatures, so the operator can stop and correct before the defect forms.
Planned · vision + audio (Wave 2)Will capture a senior welder's corrections in their final working months and learn the unwritten rules – the torch dance for galvanized steel, when to skip a tack, the sound of a cold start – so the knowledge stays after they leave.
Planned · knowledge capture (Wave 3)Will integrate natively with the welding controller: read welding parameters, suggest Technology File presets and walk teach-mode programming step-by-step on the first new part.
Planned · controller onboarding (Wave 2)Will expose defect detection to OEMs and integrators: send a frame and audio chunk, receive arc-state classification as JSON, edge or cloud, pay-per-call.
Planned · defect API (Wave 3)Verified – not promised
The cloud diagnostic loop is deployed and running. The points below are from the live system, not a pilot.
Every model answer passes a deterministic safety post-filter before it reaches the user. Imperative equipment commands are masked; unsourced claims are flagged for senior review.
Deployed · INV-001 · 8/8 safety testsThe agent has no equipment-control tools – it issues recommendations only. Even a hallucinated command is stripped before it reaches the operator.
By design · INV-002All cloud processing runs in AWS eu-central-1 (Frankfurt). Raw audio is off by default; only derived features ever leave the device.
Live · INV-016 / INV-006Ask in plain language and get a diagnosis with cited sources and safetyChecked=true – end-to-end in about 3.9 seconds on the deployed system.
Live · E2E ~3.9 sNot built yet
Named so the gap between what runs today and what is still a plan stays visible.
Moving from frame-level classification to segmenting the pool and the bead, so an advisory can point at geometry rather than at a label. This is what makes the difference between "porosity risk" and a reason.
Planned · vision model (Wave 2)Vision, acoustics, arc current and torch angle on one monotonic clock, with the degraded paths already tested in the mock stack. Sync is the precondition for fusion: if it fails, the system drops to vision-only rather than guessing.
Planned · INV-005 · on real JetsonThe avoided-reject assumption both economic models rest on is unearned until a cell is measured before and after. Two pilots, baselines taken first, criteria agreed in writing — that is the falsification plan, not a sales target.
Planned · where the models get testedDesigned for EN ISO 3834, EN ISO 5817 and EN 15085. ARCMENTOR is advisory: weld qualification and certification remain the responsibility of the shop's certified IWE/IWS.
Funding fit
EIT RawMaterials sub-theme: digital solutions for monitoring, analytics and modelling. Every claim below is checkable against the repository.
A two-loop edge/cloud architecture with a deterministic safety layer between the model and the operator. Multimodal arc vision and acoustic features. Material-passport data is derived from the welding process itself — not reconstructed afterwards from ERP exports.
A double bind. The European Welding Federation projects a shortfall of 330,000 qualified welders by 2027. At the same time ESPR and CBAM will demand product-level material and process data that Tier-2 fabricators do not physically hold.
Tier-2 fabricators in Central and Eastern Europe, where EU reshoring lands first. The CBAM definitive regime starts in January 2026; ESPR prioritises iron and steel, then aluminium. No product-specific act is in force yet — which is exactly the window to build the data layer.
A 14-sprint roadmap with six gates. The cloud diagnostic loop is live: end-to-end ~3.9 s warm, every response safety-checked. 16 Lambdas, 7 DynamoDB tables, cosign-signed immutable container images. The edge stack runs end-to-end today as a CPU mock; real Jetson hardware is Sprint 9.
A solo founder, stated plainly. Human-in-the-loop senior review is built into the product, not promised: novel cases route to a certified reviewer before they reach an operator. An IWE/IWS domain advisor is being brought on; the legal entity is in registration.
Rebuild Ukraine
Ukraine's reconstruction is, materially, a welding problem: bridges, rail, energy infrastructure, structural steel. The welders who hold that knowledge are mobilised, displaced, or retiring. Every tonne of steel not scrapped at the arc is a tonne available for reconstruction.
ARCMENTOR is pre-pilot. There are no pilot cells, no customer references, and no measured reconstruction impact yet. What exists is the layer that makes that impact measurable.
New European Bauhaus
Less scrap at the arc, process evidence that a material passport can be built on, and closed material loops. The savings model and the field-coverage walk are both published and checkable.
Skilled welding is gatekept by tacit apprenticeship. A co-pilot that watches the arc and prompts in the operator's own language lowers that barrier — for newcomers without a mentor, for veterans and displaced workers retraining into industrial work.
Human-centric Industry 5.0. The system recommends and never commands the welding equipment — that is an architectural invariant, not a setting. Advisory only · the operator stays in control.
We do not automate the welder away. We make the craft learnable again.
Beachhead
Not "manufacturing". One buyer, one channel, one countable market — narrow enough that being wrong about it shows up quickly.
At a Tier-2 fabricator in Poland, Romania or Ukraine running a handful of robotic cells. The person who owns the reject rate and the delivery date — not the CTO, and not the person who signs software licences.
Robot-cell integrators already have the access and the trust: they installed the cell, they return for service, and they are asked what to bolt onto it next. ARCMENTOR ships as an addition to a cell they built, not as a rip-and-replace.
The beachhead is counted in robotic welding cells reachable through those integrators, not in a top-down share of a global market figure. Small on purpose: it is a number that can be checked cell by cell.
Why This Is Hard To Copy
A camera that flags porosity after the fact reports a loss that already happened. The value sits in the seconds before the arc reaches that condition again — and that is a different system, not a better classifier.
Inspection tells the operator what to scrap. An advisory loop that runs while the arc is burning tells them what to change — which is why the latency budget and the offline requirement exist at all.
Novel cases route to a certified welding engineer before they reach an operator, and their verdicts become the labels. The dataset is a by-product of the review loop customers are paying for — it is not on the open internet and it is not for sale.
Every advised weld leaves a machine-readable trace — parameters, geometry, operator, serial number — grounded in ISO 3834, EN 15085 and EN ISO 5817. That record is what a product passport will need, and a defect classifier does not produce it.
How We Start
Three steps, one cell at a time. We publish no price list: the pilot sets the baseline, the success gate decides, and commercial terms follow the savings we can actually show.
We fit the kit to a cell you already run: camera, mic, headset, edge box. One operator, one cell, a fixed pilot window. Your data stays yours.
At the end of the pilot we compare against the baseline measured on day one. Criteria are agreed in writing before the pilot starts, so you decide on numbers rather than on a demo.
Only after a go: we extend to further cells, and the price is tied to confirmed savings rather than to a licence list. Terms are agreed per site.
Live Demo
Ask the agent to analyze a weld parameter set, suggest a coach prompt, or draft a WPS for a sample joint.
The Ask
Stated plainly, because a vague ask is an ask nobody can act on.
Steel or aluminium fabrication in CEE, one robotic cell each, willing to have a baseline taken before anything is installed. The pilot is where the avoided-reject assumption is earned or falsified.
Robot-cell integrators in the region, and access to the EIT RawMaterials industry network. The channel matters more than the pitch: integrators are how a cell-level product reaches cells at all.
Cloud diagnostic loop live, edge stack running as a CPU mock with the safety and degradation paths tested, sensor hardware acquired and the port map complete. Two lab cells acquired — not yet commissioned.
Get In Touch
Have a robotic welding cell with a training bottleneck? Let's talk about a 14-day pilot in your shop.
Join the early-access list to help shape the pilot – no commitment.
By emailing us you accept our Privacy Policy.
Agentic AI · Architecture
A supervisor agent orchestrates specialized worker agents that fuse vision (Lucid Triton IMX490 HDR), audio (miniDSP UMIK-1) and historical WPS records. Edge inference handles real-time coaching; the cloud handles nightly retraining; Human-in-the-Loop senior welders confirm novel cases before the model promotes them to production.
Compliance Framework
Tiered responsibility for welding quality assurance – from the operator at the cell, through the shop's IWE/IWS, up to EU directives. ARCMENTOR is the audit trail that connects them.