01
Foundation models grew up.
Multimodal models can now extract structured procedures from video and audio at industrial fidelity. What needed a forty person annotation team in 2023 is now a prompt and a pipeline.
Memory for the plant floor
Floorstack captures expert interventions directly on the line. Speech, video, context and asset data become structured precedent for the next similar stoppage.
Foundation models can now extract reliable procedures from speech and video. At the same time, the expertise of the best operators is leaving the plant.
01 / 06 · Plant alert
Stoppage reported.
01 / 07 · Manifest
Industry runs on tacit knowledge. The kind of judgment a reliability engineer earns over twenty eight years and that vanishes the day they retire. The day the last good diagnostician on a line walks out, the line forgets.
Floorstack treats that knowledge as infrastructure. We capture the intervention, not the documentation. We index it against the asset, not against an audit. And we put it back into the operator's hands the next time the same symptom returns, on a different shift, in a different plant, in a different country.
02 / 07 · The Problem
lost annually to unplanned downtime across the world’s 500 largest manufacturers. Eleven percent of their combined revenue.
Source: Siemens, True Cost of Downtime 2024
of employers cite skills gaps as the number one barrier to transformation. Manufacturing leads in reskilling need through 2030.
Source: World Economic Forum, Future of Jobs Report 2025
01
An unusual problem arises.
A symptom no junior has seen before. A senior intervenes.
02
It gets documented too late, if at all.
By the time someone writes it up, the language is sanitized for audit and the diagnostic reasoning is gone.
03
The next person starts from scratch.
On the next shift, in the next plant, in the next quarter. The same hypothesis tree is rebuilt by hand.
03 / 07 · Approach
Four steps. Each one runs in the background until an expert begins work on a defined event.
First person video and ambient audio, time synced with asset and process context. Triggered, not continuous: capture starts when work begins on a fault, deviation, or non standard intervention. Ends when the event ends.
We convert the captured episode into structured data: symptom seen, hypothesis stated, action taken, equipment referenced, outcome observed. Vision and audio extraction, validated by the operator.
Every record is indexed against your asset hierarchy, so the next operator can retrieve a precedent by symptom, by equipment, by alarm code, on the line they’re standing on.
When a similar event recurs, the relevant precedent is surfaced in seconds. The same corpus later grounds audio prompted guidance for the operator and any AI agent acting inside the plant.
04 / 07 · Industries
The platform is horizontal. The vocabulary is not. Each industry gets its own asset hierarchy, its own evidence standard, and its own retrieval grammar.
Turnarounds, LOPC events, root cause failure analysis. Designed for IECEx and ATEX environments and air gapped operation.
CNC, EDM, grinding, measurement. Captures setup expertise, SPC interventions, and the kind of fixturing judgment that lives in the head of one setter.
Validated processes under ISO 13485 and FDA QMSR. Capture and replay aligned with CAPA, deviation handling, and 21 CFR Part 11 audit trail expectations.
Body shop robotics, paint, assembly. Andon triggered capture, Yokoten style replay across sister plants. Works council aware capture design.
05 / 07 · Why now
01
Multimodal models can now extract structured procedures from video and audio at industrial fidelity. What needed a forty person annotation team in 2023 is now a prompt and a pipeline.
02
RealWear and Vuzix now ship intrinsically safe IECEx Zone 1 head mounted devices in production. Hands free capture on the line moved from pilot to deployment this year.
03
An entire generation of senior operators is retiring at once across industries. The window to capture their judgment without asking them to write it down is years, not decades.
06 / 07 · Company
A founding team out of TUM, with research at Stanford, Cambridge, MIT, IBM and Fraunhofer behind us. Advised by operators who’ve built and scaled industrial software before.

Co founder
Management and CS at TUM. Research on industrial energy flexibility at Stanford. AI implementation work at Langdock (Trumpf, Würth).

Co founder
Management and CS at TUM. Research at IFM Cambridge and MIT on AI in manufacturing. Built an enterprise grade AI system inside the TDK incubator.

Co founder
Informatics and Physics at TUM. ICML published machine learning work. Currently VP at TUM.ai. Quantum ML research at Fraunhofer and IBM.
Advisors
Thomas Kirchner
Co founder and former CEO, ProGlove
Prof. Thomas Bohné
Founder, Cyber Human Lab, University of Cambridge
Dr. Slavomir Tadeja
Postdoc, MIT LEAP Group, workforce development
Research and training environments
07 / 07 · Engage
We start with a two to three day on site engagement. Shadowing five to eight expert operators during live maintenance and fault resolution. Full NDA. Safety induction. Zero operational disruption.
What you get
What you commit