Technology

Physical intelligence

for modern factories.

Factories are where the physical world gets made,

but most of what happens there is still invisible.

Almetra develops AI systems that learn from real shopfloor behaviour turning factory video into structured process understanding, operational insight, and automation readiness.

Manufacturing work unfolds over minutes, across people, tools, parts, machines, stations, and process variants. Our systems combine temporal video understanding, process mining, typed process graphs, large teacher models, edge-deployed students, and human process expertise.

THE TECHNICAL PIPELINE

From pixels to action

six transformation layers.

LAYER 01

Pixels

Raw video

LAYER 02

Entities

People · tools · parts

LAYER 03

Actions

Steps · motions · waits

LAYER 04

Process Events

Cycles · bottlenecks

LAYER 05

Process Graph

Typed · auditable

LAYER 06

Insight & Action

CI · automation

Vision AI

Detection

Temporal model

Event extraction

Graph construction

Improvement & robotics

TECHNICAL STACK

Six transformation layers

from raw video to operational decisions.

LAYER

WHAT IT DOES

TECHNICAL SIGNAL

Tech. SIGNAL

Pixels → Entities

Video becomes robust representations of people, hands, tools, parts, workstations, and scene regions.

Messy real-world perception under occlusion, lighting variation, and factory noise.

Pixels → Entities

Video becomes robust representations of people, hands, tools, parts, workstations, and scene regions.

Messy real-world perception under occlusion, lighting variation, and factory noise.

Entities → Actions

Temporal-aware models infer steps, motions, interactions, waiting, rework, and deviations over minutes.

Long-horizon video understanding beyond frame-level detection.

Entities → Actions

Temporal-aware models infer steps, motions, interactions, waiting, rework, and deviations over minutes.

Long-horizon video understanding beyond frame-level detection.

Actions → Events

Activity mapped into cycle starts, value-add work, idle time, bottlenecks, and standard-work deviations.

ML outputs become manufacturing events instead of labels for their own sake.

Actions → Events

Activity mapped into cycle starts, value-add work, idle time, bottlenecks, and standard-work deviations.

ML outputs become manufacturing events instead of labels for their own sake.

Events → Process Graph

A typed graph represents states, transitions, variants, dependencies, loops, SOP constraints, and failure modes.

Process intelligence, ontology design, auditability, and operational reasoning.

Events → Process Graph

A typed graph represents states, transitions, variants, dependencies, loops, SOP constraints, and failure modes.

Process intelligence, ontology design, auditability, and operational reasoning.

Graph → Insight

The system explains where time is lost, why output varies, which steps constrain capacity, and what to improve.

AI that directly affects production decisions.

Graph → Insight

The system explains where time is lost, why output varies, which steps constrain capacity, and what to improve.

AI that directly affects production decisions.

Insight → Automation

Process context identifies automation candidates and gives robotic systems the task context they need.

The bridge from observation to robotic execution.

Insight → Automation

Process context identifies automation candidates and gives robotic systems the task context they need.

The bridge from observation to robotic execution.

TECHNICAL CHALLENGES

Teaching AI to understand

how factories work.

  • LONG-HORIZON VIDEO

    Seeing work, not frames

    Manufacturing processes unfold over minutes with temporal dependencies, action segmentation, and process state that frame-level detection cannot capture.

  • TEACHER-STUDENT DISTILLATION

    Compressing large models for edge deployment

    Transferring the capabilities of large temporal teacher models into compact student models that run close to the production line, under real factory latency and bandwidth constraints.

  • VIDEO-TO-PROCESS GRAPHS

    Reliable process events from noisy shopfloor video

    Converting raw factory video into typed process graphs with auditable transitions, dependencies, and failure modes, structured enough for operational reasoning and automation decisions.

  • CROSS-FACTORY ADAPTATION

    Adapting across factories with limited labels

    Generalising across factories, workstations, operators, camera viewpoints, and product variants without requiring exhaustive annotation at each new deployment.

  • CYCLE MATERIALISATION

    Deriving cycles from dense workstep streams

    Extracting clean cycles and per-product bookkeeping from continuous, dense workstep streams to replace brittle counting heuristics with reliable structured event data.

  • AUTOMATION READINESS

    Which tasks are worth automating, and when

    Determining which human tasks are economically meaningful, technically feasible, and operationally safe to automate, grounded in real process data.

  • LONG-HORIZON VIDEO

    Seeing work, not frames

    Manufacturing processes unfold over minutes with temporal dependencies, action segmentation, and process state that frame-level detection cannot capture.

  • TEACHER-STUDENT DISTILLATION

    Compressing large models for edge deployment

    Transferring the capabilities of large temporal teacher models into compact student models that run close to the production line, under real factory latency and bandwidth constraints.

  • VIDEO-TO-PROCESS GRAPHS

    Reliable process events from noisy shopfloor video

    Converting raw factory video into typed process graphs with auditable transitions, dependencies, and failure modes, structured enough for operational reasoning and automation decisions.

  • CROSS-FACTORY ADAPTATION

    Adapting across factories with limited labels

    Generalising across factories, workstations, operators, camera viewpoints, and product variants without requiring exhaustive annotation at each new deployment.

  • CYCLE MATERIALISATION

    Deriving cycles from dense workstep streams

    Extracting clean cycles and per-product bookkeeping from continuous, dense workstep streams to replace brittle counting heuristics with reliable structured event data.

  • AUTOMATION READINESS

    Which tasks are worth automating, and when

    Determining which human tasks are economically meaningful, technically feasible, and operationally safe to automate, grounded in real process data.

INFRASTRUCTURE & RESEARCH

Built on production data,  

serious systems, and frontier research.

TRAINING INFRASTRUCTURE

Blackwell-class compute for temporal video models

Self-hosted high-performance training infrastructure for training and adapting large temporal models on factory video at scale.

ACADEMIC RESEARCH

Computer vision and robotics research depth

Academic advisors and collaborators help connect frontier research with production deployment challenges.

ROBOTICS & EMBODIED AI

From process understanding to robotic skills

We work with robotic arms and automation testbeds to explore embodied reasoning, vision-language-action models, and skill learning grounded in real factory context.

DATA ADVANTAGE

Real factory data at scale

Almetra observes manufacturing work across deployed production environments creating a rare dataset for temporal video understanding, process graph construction, and automation-readiness analysis.

Temporal video, process graphs, edge AI, and robotics

See what's happening on your manual line

deployed in real factories.

We are building the physical intelligence layer for manufacturing. If long-horizon video, process graphs, teacher-student distillation, edge AI, or robotics with real production context sounds like the right frontier, we’d like to hear from you.

Request a demo or start with a factory assessment. Most teams see their first real-time insights within two weeks.

Watch recent customer story