SHMEL — The human-skill data engine for dangerous trades

What everyone else is missing

Everyone is racing for the wrong data.

The easy data — the whole crowd

Warehouse, kitchen, tabletop tasks. Cheap to collect, so everyone collects them. Prices fell 3× in two years — a race to zero against $70M war-chests and low-cost gig networks.

The data capital can't buy — us

Dangerous, high-skill work. You can't scrape it. You can't cheaply teleoperate it. You reach it only through trust and access — a network of skilled workers who let you record them. That's the corpus we own.

What we do

Capture → validate → own.

We fit skilled workers with wearable capture, turn their demonstrations into training-ready data, and keep the corpus — so it sells again and again.

Step 01

Capture

Skilled workers wear an egocentric rig and do the job — real demonstrations of hazardous, dexterous work, narrated and consented.

POV · on-site · hands-free

Step 02

Validate — our IP

We turn raw footage into action-labeled, consent-verified, co-training-ready data with a defensible chain of title. The validation layer is the moat, not the camera.

Action labels · provenance

Step 03

Own & re-license

Exclusive datasets and eval sets to VLA (vision-language-action robot) labs and OEMs — and we keep the corpus. Non-exclusive by default, so it compounds.

Owned · re-licensable

How we capture

We capture the whole signal — not just video.

A welder's skill lives in far more than a camera feed. Our wearable rigs record every channel a robot needs to actually learn the job.

01

Vision

Head-mounted POV of the exact task — the robot's-eye view.

02

Sound

Arc tone, spatter & tool audio that signal a clean weld.

03

Touch

Force & grip from sensor gloves — the dexterity robots miss.

04

Voice

The worker narrates decisions, mistakes & recoveries.

05

Motion

Hand & body pose, torch angle, travel speed.

Gear today: DJI head-mounts, sensor-equipped glasses & tactile gloves — rented to workers. Our own capture hardware is on the roadmap.

Why welding first

The perfect beachhead — and the buyers already exist.

$5–6B global welding-robotics market by 2030, with adaptive welding — the kind robots still fail at — the fast-growing segment we serve.

Un-automatable

Adaptive & repair welds

Not the fixed-path robots that already exist — the variable, in-the-field welding that still needs real human skill.

Dangerous → in demand

Safe to capture, urgent to automate

Fumes, arc and burns drive the automation pull — but it's a shop floor, not a reactor, so we can put rigs on welders today.

Buyers exist now

Path, Novarc & a 300k+ shortage

Real companies are building AI welding right now, and a deep welder shortage makes the demand undeniable.

The niche map

One capture engine. Every dangerous trade.

We conquer a hazardous vertical, own its corpus, then stack the next. 50+ funded robotics companies already build in these niches — ~$2.5B raised into them — and no one owns the human-skill data they all need.

~11 buyers

Beachhead

Welding & fabrication

The adaptive welding robots still can't do. Anchor buyers: Path Robotics, Novarc.

~15 buyers

Open

Energy trades

Wind-turbine techs, high-voltage, lineman — heights and electrical, a booming sector.

~8 buyers

Open

Foundry & metal fab

Grinding, casting, forging — extreme heat, high-skill, welding-adjacent expansion.

~12 buyers

Open

Hazmat & nuclear decon

Dangerous, low-glamour work robots want — and no data engine is capturing it.

~7 buyers

Open

Meat & food processing

Butchery, deboning, trimming — pure dexterity, un-hunted, and robots keep failing at it.

Skip · crowded

Warehouse & construction

Already being nibbled (Cortex, Canvas, Hardhat). We avoid the commoditizing crowd.

Then: nuclear & chemical · offshore & deep-sea · semiconductor · shipbuilding · steelworks · wastewater — the same engine, one dangerous trade at a time.

A two-sided engine

Robotics companies get the data. Welders get paid — for life.

For robotics companies

Cleaned, action-labeled, co-training-ready datasets — including the failure & recovery data teams need for robustness that no one else captures. Reusable, on demand, with a chain of title they can defend.

For welders & shops

Earn royalties every time your data is re-used, get paid weekly, and own your chain of title. We recruit through fab shops, trade schools & creator channels — a supply network that compounds.

Why we win — and why they can't follow

Three moats, stated plainly.

We own the dangerous long tail

The hardest data to get, the most valuable to automate. You can't scrape it, you can't cheaply teleoperate it, and no incumbent is built to reach these workers.

We're independent

Meta owns 49% of Scale AI — the top labs won't feed a Meta-affiliated vendor. We're the neutral supplier, with provenance-grade data lineage they can defend in court.

We own the asset, not the labor

Everyone else resells labeled hours — a deflating commodity. We own a compounding corpus, and workers own a share. Our margin rises while theirs falls.

How it works