OUR MISSION
The physical world
deserves the same data discipline
as the digital one.
Digital workflows are logged, analysed, and optimised as a matter of course. Physical operations, the shop floor, the assembly line, the factory, still run largely on assumption, periodic observation, and experience that walks out the door when people leave.
Almetra exists to close that gap. Not with more dashboards, but with a genuine intelligence layer that captures how production actually works and turns that knowledge into improvement, automation readiness, and production blueprints.
WHO WE ARE
Not AI people who discovered factories.
Factory people who know AI.
Most industrial AI companies start in a lab and work outward. We started on the shop floor and worked inward, with operators, plant managers, and CI teams who needed answers that no existing system could give them.
That difference shows up in what we build. The edge constraints, the union approvals, the IT/OT realities, the fact that a camera going offline for six hours means a line goes dark, these are not edge cases we design around. They are the centre of our architecture.
WE ARE NOT
A wrapper around a foundation model. A dashboard company with cameras bolted on. An AI lab that has never seen a factory floor at shift change.
WE ARE
Building our own temporal models, trained on thousands of hours of real factory video, deployed on-device under real production constraints, and generating intelligence that directly changes how factories operate.
THE TECHNICAL FRONTIER
Hard problems, real constraints,
production scale.
We build our own models. We do not wrap APIs or plug in off-the-shelf vision libraries. The problems we work on are frontier, and they are only solvable because we have access to real factory data at scale.
TEMPORAL VIDEO UNDERSTANDING
Minutes, not frames
Manufacturing work unfolds over minutes — operators, tools, parts, variants, machine states, rework, and handovers. Our models capture this temporal structure, not just what is visible in a single frame.
TEACHER-STUDENT DISTILLATION
Large models → compact edge deployment
We train large temporal-aware teacher models on factory video and distil their capabilities into per-station student models that run on-device, close to the line, under real latency and bandwidth constraints.
VIDEO-TO-PROCESS GRAPHS
Structure from noise
Converting noisy shopfloor video into typed process graphs — with auditable states, transitions, variants, dependencies, and failure modes — structured enough to drive operational decisions and support robotic task context.
AGENTIC DATA CURATION
Station onboarding as an agent problem
Raw video and a station description go in. Agents handle review, labelling, reranking, and materialisation — producing a fine-tuned per-station VLM and structured cycle data, with auditable provenance at every step.
CROSS-FACTORY ADAPTATION
Generalising with limited labels
Adapting across factories, workstations, operators, camera angles, and product variants without requiring exhaustive annotation at each deployment — one of the hardest open problems in applied video understanding.
AUTOMATION READINESS
Which tasks are worth automating
Determining which human tasks are economically meaningful, technically feasible, and operationally safe to automate — grounded in real process data from real factories, not simulations or lab environments.
Open Roles
Work on a hard problem that matters
in production, not a lab.
We are a small, senior team working across AI and computer vision, backend and data infrastructure, product and design, and customer-facing deployment. Berlin-based, international, direct.
Experience in manufacturing is not required. Curiosity about it is — and so is the kind of rigour that comes from caring whether the thing you built actually works when the shift starts at 6am.
Don’t see a role that fits?
We’re always looking for exceptional people.
Send us a message and explain why you're a great fit.
System status · OperationaL
v 2026.5