Skip to content
Blueprint · Manufacturing

Battery quality deviation. From cell test to pack decision.

A worked example: a capacity deviation in end-of-line cell testing, investigated across live plant events, genealogy, process data and specifications. Agents do the legwork. Engineers make the call.

Manufacturing
The question

Why did capacity drop on line B this week, and which packs are affected?

Inputs

What the agents read.

Cell test results
End-of-line capacity and impedance by cell
Genealogy
Material lot → cell → module → pack
Process data
Formation and ageing parameters by batch
Live plant events
Machine states, alarms and counts by station
Specifications
Current limits, SOPs and their revision history
Run

What happens, step by step. Each step is logged with its sources.

  1. 01

    Scope the deviation

    Find the affected cells, the batches they share and when the drop started.

  2. 02

    Check the process

    Compare formation and ageing parameters for affected and unaffected batches.

  3. 03

    Trace the material

    Follow material lots through genealogy to every module and pack they reached.

  4. 04

    Check the rules

    Read the current specification and SOP revision for this cell type and step.

  5. 05

    Draft the disposition

    Propose hold, retest or release per pack, with the evidence attached.

Agent Hive · Run 4107
Why did capacity drop on line B this week, and which packs are affected?
  1. Queriedmes.cell_tests · 18,240 cells1.8s
  2. Comparedformation log, chambers 3 and 41.1s
  3. Tracedlot M-114 → 14 packs0.9s
  4. ReadSOP formation rev 7 · limit 3.2 Ah0.6s
  5. Draftedhold proposal with 5 sources0.4s
Answer4 sources
−2.8%3.11 Ah
Median capacity, line B, week 41

The drop sits in batch B-77 on formation chamber 3, all from material lot M-114. 14 packs are affected; hold proposed for review.

cell_testsformation_loggenealogySOP formation r7
Ask a follow-up…
Approvals

Where people decide.

  • Needs you

    Hold scope

    The quality engineer confirms which packs are held before anything changes in the MES.

  • Needs you

    Corrective action

    Process changes are proposed to the line owner, never applied directly.

  • Needs you

    Knowledge changes

    New entities, relations and rules the agents suggest enter the plant graph only after an engineer accepts them.

Outputs

What you get back.

Investigation report
Hypotheses ranked, each with linked evidence
Affected packs
Genealogy path and recommended disposition per pack
Run trace
Every query, document and graph version the agents used
Data sources

Live events and engineering documents, in the same picture.

Both are mapped to one plant model, so an alarm, a process step and the right procedure can be connected. Sources stay where they are; connectors are read-only.

Fig. 1 · Connected sources
Sources stay in place; Agent Hive reads them through read-only connectors and builds one shared model. Logos are trademarks of their respective owners.
  • Live operational events

    Machine states, values, alarms and counts, mapped to equipment and process steps.

  • Engineering documents

    SOPs, recipes, datasheets and specifications, read for equipment, limits and applicability.

  • Genealogy

    Material, lot, cell, module and pack history from the MES and ERP.

Agent workflow

Specialists work in parallel. The hold waits for a person.

The run fans out to process, material and specification agents, joins their findings and stops at the hold. Nothing changes in the MES until the quality engineer signs off.

Fig. 2 · Workflow runIllustrative data
The run as the platform draws it: tool steps, specialist agents, joins, branches and the approval it is waiting on. Drag to pan; Ctrl or ⌘ + wheel to zoom.
Knowledge model

One plant graph. Agents propose, engineers approve.

Agents suggest new entities, relations and rules as they read the sources. Engineers review every change before it becomes part of the graph, and each answer stays tied to the graph version that produced it.

Fig. 3 · Knowledge modelIllustrative data
The plant graph around batch B-77: process steps, equipment, specifications, SOPs and genealogy, each node linked to its source table or document.
  • 01

    Suggest

    Agents read the sources and suggest entities, relations and rules.

  • 02

    Review

    Process engineers accept, edit or reject each proposed change.

  • 03

    Add to the graph

    Approved changes are versioned and linked to their sources.

  • 04

    Find what is missing

    Uncertain links and unanswered questions go back for review.

Fig. 4 · Approved changes onlyIllustrative data

Ontology evolution. How the structure changed over 4 proposals: v1: baseline with Thing, Document, Report, Metric, Party, Vendor, Supplier. P-1 (add, from steward): Add class Definition ⊑ Metric, applied, giving v2. Diff: + Definition subClassOf Metric. P-2 (merge, from resolver): Merge Vendor into Supplier, applied, giving v3. Diff: − Vendor; + Supplier altLabel “Vendor”. P-3 (add, from extractor): Add class Misc ⊑ Document, rejected; v3 unchanged. Diff: + Misc subClassOf Document; ✗ rejected: no defining property. P-4 (deprecate, from steward): Deprecate Report, applied, giving v4. Diff: ~ Report owl:deprecated true; + replaced by Document. Selected: P-4 (deprecate, from steward): Deprecate Report, applied, giving v4. Diff: ~ Report owl:deprecated true; + replaced by Document.

The graph changes through proposals: agents suggest, engineers accept or reject, and every version is kept.
Evidence and provenance

Every pack has a path. The hold follows it.

Genealogy decides the hold scope: the material lot the evidence points to, traced through cells and modules to the packs it reached.

Fig. 5 · Genealogy and hold scopeIllustrative data
MATERIAL LOTFORMATION BATCHCELLSMODULESPACKSLot M-113cathodeLot M-114cathodeLot M-115anodeBatch B-76chamber 4Batch B-77chamber 3Batch B-78chamber 42,380 cellsin spec2,402 cells−2.8% capacity2,395 cellsin specModules 0410-0438Modules 0439-046628 modulesModules 0467-0495Packs 2201-2214Packs 2215-2228hold proposedPacks 2229-2242
Material lot M-114 traced through batch B-77 and its cells to the modules and packs it reached. Only the violet path is proposed for hold.
Question types

Test it with the questions engineers actually ask.

  • Find a fact

    What is the current specification?

    Return the right setpoint, limit or procedure for this product and process.

  • Show the evidence

    Where did this answer come from?

    Link the events, documents, lots and timestamps behind it.

  • Investigate a cause

    What is the most likely cause?

    Compare causes across process areas and show the supporting evidence.

  • Check a change

    What else changes if we do this?

    Find the affected specifications, procedures, limits and downstream steps.

Evaluation

Every version is graded. Before it reaches the line.

A question set agreed with plant engineers checks that answers are correct, complete and backed by the right evidence. Each reviewed answer joins the set.

Fig. 6 · Question set by versionIllustrative data
Evaluation results by question and version
Questionv1v2v3v4
What is the formation voltage limit for cell type A?Find a fact✓✓✓✓
Which SOP revision applies to ageing on line B?Find a fact✗✓✓✓
Which material lots fed the failed cells?Show the evidence~✓✓✓
Which packs contain cells from lot M-114?Show the evidence✗✗✓✓
Why did capacity drop on line B in week 41?Investigate a cause✗~✗✓
Is the drop on one formation chamber or all?Investigate a cause~✓✓✓
If ageing time drops by 4 h, which limits change?Check a change✗✗~✓
Which downstream tests depend on the impedance limit?Check a change✓~✓✓
v4: 8 of 8 pass · release needs all 8
Each row is a question agreed with the team; each column a version of the workflow. The candidate ships only if it holds the release bar.
  • 01

    Engineer reviews the answer

    Correct, partly right or wrong, with a note.

  • 02

    The question joins the set

    Reviewed answers become repeatable cases.

  • 03

    Each new version is checked

    Prompts, tools and graph changes rerun the set.

  • 04

    Only passing versions ship

    A named owner approves the release.

Outcomes

What changes on the line. Measured on the question set and the run log.

To a scoped hold
45 min
from two days of manual pulls
Held, not a week
14 packs
genealogy-based hold scope
Answers with sources
100%
every claim links to evidence
Eval pass, v4
8 of 8
release bar for the line
Illustrative data
Deployment

Keep the same graph, whichever model setup you choose.

Models can run inside the plant or through a controlled gateway. Plant context and access controls stay the same.

Run everything inside your network.

Fully air-gapped

The knowledge graph, agents and models run locally. Updates arrive as signed offline packages and no data leaves the site.

  • Local model serving
  • Read-only connectors
  • Your SSO, roles and audit

Keep the data local, use approved models.

Hybrid model access

A policy gateway controls what is sent, which model receives it and what comes back into the workflow.

  • Model choice per agent role
  • One controlled egress path
  • Graph and sources stay local
Fig. 7 · Air-gapped perimeter
Air-gapped deploymentEverything runs inside the customer’s perimeter with no outbound internet: users, the agent runtime, local models, the knowledge graph, the audit log and read-only connectors to on-prem systems. Updates come in as signed offline bundles. Point at a part to see what it holds and what is blocked.OUTSIDEReleasesigned bundlevX.Y · checksumcarried on offline mediano network linkINTERNEThosted model APIsCUSTOMER PERIMETERNO EGRESSUsersSSO · RBACcustomer identityAGENT HIVE RUNTIMEAgents · workflowsApproval gatesPoliciesEvals · tracesLocal modelsGPU servinginside the siteKnowledge graphdocuments + systemsstored locallyAudit logwho · what · whenevery approvalON-PREM SYSTEMSREAD-ONLY CONNECTORSERPData lakeDocumentsagents read through approved, read-only accessREADUpdate importverify signatureadmin approvesAPPROVED UPGRADEno outbounddata, models, graph and audit never leave the line

Point at a part no egress · signed bundles in · data stays inside

Fully air-gapped: graph, agents, models and traces inside the plant network. Updates arrive as signed offline packages.

Run this blueprint on your data. We set up the first one with you.