KnowledgeRelay
Expert knowledge turned into structured work instructions before the experts retire.
- Category
- Knowledge capture
- Model
- On-prem or SaaS
- Core platform
- On-prem extraction pipeline
- Status
- IN DEVELOPMENT
01 Problem
Manufacturers are losing decades of process knowledge to retirements, and it leaves in the heads of the people who walk out. The manuals are outdated, the notes are handwritten, and the real procedure exists only as what the expert actually does.
The cost arrives after the retirement party: scrap rates climb and nobody remembers why the third step matters.
02 Specification
This is the specification we are building toward. The status column shows what exists today.
- Extraction from manuals, notes and expert videoIN PROGRESS
Knowledge is extracted from old manuals, handwritten notes and recorded expert walkthroughs.
IN PROGRESS - Structured work instructionsPLANNED
Extracted knowledge becomes structured, step-by-step work instructions instead of a searchable pile of documents.
PLANNED - Cross-language outputPLANNED
Instructions are delivered in the operator’s language, whatever language the source material was in.
PLANNED - Offline and on-prem operationPLANNED
The system runs offline and on premises, so process knowledge never leaves the building.
PLANNED - Confidence scoring with human reviewPLANNED
Every extracted instruction carries a confidence score and passes through human review before an operator sees it.
PLANNED - Equipment-specific knowledge graphsPLANNED
Knowledge is organized per machine and per process, so instructions carry their context.
PLANNED - Searchable institutional memoryPLANNED
Everything captured becomes searchable, so the answer to why the third step matters survives the person who knew it.
PLANNED - Update workflowPLANNED
When a process changes, instructions are revised through a controlled update workflow instead of drifting out of date.
PLANNED
Updated as capabilities ship. BUILT means working and validated; nothing gets that label early.
03 System
- 01Documents and expert video
- 02Domain extraction
- 03Structured instructions
- 04Review loop
- 05Operator delivery
04 Engineering
Extraction pipeline
KnowledgeRelay combines document understanding for degraded scans and handwriting with video understanding for recorded expert walkthroughs, mapped into equipment-specific knowledge graphs. Extraction models run entirely on premises, and every extracted step carries a confidence score.
Review and delivery
Nothing reaches an operator unreviewed: low-confidence extractions route to a human review queue by design, and reviewers correct in place, which tunes the system on the plant’s own material. Delivery is structured work instructions in the operator’s language.
Deployment model
The system runs offline and on premises, so process knowledge, drawings and video never leave the building. A hosted option exists for plants that want one, but the design assumption is that most will not.
05 Why this is hard
The genuine problem is extraction accuracy on noisy domain documents and video, in multiple languages, wrapped in review workflows operators will actually accept. A faded decades-old manual, a machinist’s shorthand and a walkthrough filmed on a phone defeat generic extraction, and a system that is confidently wrong gets rejected on the floor within a week. The engineering work is domain-tuned extraction that knows what it does not know and a review loop that earns operator trust instead of assuming it.
06 Design constraints
- Runs offline and on premises
- Knowledge never leaves the building
- Nothing reaches an operator unreviewed
- Every extraction carries a confidence score
- Delivered in the operator’s language
07 Design partner program
Design partner program.
We are looking for a small number of design and pilot partners. A partner gets the problem solved for their operation first, direct influence on the specification above, and pilot pricing. In exchange we get reality: real constraints, real data, real operators. No external partner is engaged on this product yet.