Skip to content
PROOF // Continuous-Process Industrial Manufacturing Multi-line food & beverage processing plant · IoT-instrumented production floor · continuous-batch operations
← All Proof

CHAPTER 03 · BUILD

Agents Shipped Inside the Existing Stack

The first system, and the one right behind it, shipped inside the plant's existing historian and CMMS — no new sensors, no new platform. Each one is real and running.


THE AGENTS · 3 FOR THIS WORKFLOW

This plant needed two. The count isn't a package — it's however many distinct jobs the work splits into. Each agent is a separate identity with its own scoped tools and permissions.

01

Sensor Briefing & Remediation

Status Runs every shift
Impact Manual watch → reviewed brief
Team Automated, 1 human review
02

Predictive Maintenance

Status Fleet health scored continuously
Impact Reactive fixes → early work orders
Team Automated, 1 human review
03

Batch-Quality & Energy Optimization

Status Wave 2 — in build
Impact Deviation tracking + cost optimization
Team Automated, plus a solver

IT DOESN'T STOP AT A REPORT

The Agent Takes the Action

A report is one thing these agents produce. Most of what they do is take the next action inside the plant's own systems — a person approves anything that carries risk.

agent inbox — this morning
DATA VIEW

Open the maintenance work order

A bearing-drift signal drafts the work order in the CMMS, pre-filled, awaiting a tech's sign-off.

Order the replacement part

Raises the parts requisition so it's on the shelf before the planned window, not after the breakdown.

Schedule the swap

Books the fix into the next planned downtime slot instead of scrambling for one.

Log the failure pattern

Writes the confirmed pattern back to the searchable equipment history automatically.

Ran automatically

PRODUCTION TRACE

Sensor Briefing & Remediation

Every step below runs against the plant's real data — the exact sequence, in order.

live.internal — this morning
DATA VIEW

3

Briefings generated

7

Checks accepted

5

Sensor flags reviewed

2

Cross-shift notes carried forward

3 items need your judgment

Q-118 Extruder Motor A bearing-drift pattern matches a known failure mode
High
Q-119 Filler Pump 2 acoustic read at 71% confidence
Medium
Q-120 Changeover sequencing suggestion for the 2:15 PM run
Medium

Ingest

Automated

Live sensor data pulled from vibration, temperature, and acoustic feeds across every line, every few minutes — no new hardware, no new wiring.

Fan out by role

Automated

One data pull becomes three briefings — plant manager (strategic), shift supervisor (execution), maintenance tech (what to check first) — each grounded in the same numbers.

Ground every claim

Automated

Every flagged reading cites the exact sensor trace or historical pattern behind it, not a general opinion.

Supervisor review

Reviewer input

Each recommended check is one-click accept or reject. Rejections become training signal for tomorrow's briefing.

Tomorrow's diff

Automated

The next briefing opens with a diff against today's accepted checks — did it catch anything, what changed.

REAL OUTPUT · PREDICTIVE MAINTENANCE

One Example, Fully Processed

When a job calls for a written artifact, this is what the system produces — everything it finds, structured for the action it triggers next.

Line 3 Extruder Motor

Vibration + acoustic monitoring · Risk scored every 15 min · Local model, no cloud round-trip

DATA VIEW

Fleet Health Scan — Line 3

Machine Signal Risk Score Est. Remaining Life Confidence
Extruder Motor A Vibration — bearing frequency drift High ~9 days 93%
Extruder Motor B Acoustic — normal Low 120+ days 97%
Conveyor Drive 3 Vibration — normal Low 90+ days 95%
Filler Pump 2 Acoustic — cavitation signature Watch ~30 days 71%
Packaging Servo 5 Vibration — normal Low 100+ days 98%

Fleet Summary — for Maintenance Planning

5 assets scanned on Line 3. Extruder Motor A shows a bearing-frequency drift pattern consistent with early-stage failure — about 9 days of remaining life at current load. Filler Pump 2's cavitation signature is a lower-confidence read and worth a manual listen before scheduling work.

Tough Calls — Maintenance's First Look

Extruder Motor A: bearing-frequency drift matches a known failure pattern. Work order drafted, awaiting maintenance sign-off.
Filler Pump 2's cavitation read is below the 85% auto-log threshold — verify with a manual listen before scheduling work.
Conveyor Drive 3 and Packaging Servo 5 are inside normal range — no action, re-scan next cycle.

WAVE 2 — IN BUILD

Batch-Quality & Energy Optimization — What's Next

Batch-quality deviations and energy costs are tracked in two different systems that never talk to each other — a slow changeover to chase a quality spec can quietly double a line's energy cost for the run.

The agent proposes a changeover-sequencing adjustment; a constraint solver checks the math against both quality tolerance and energy cost. Nothing ships on an LLM's arithmetic — the solver's answer is the answer, the agent just explains it.

Worked example the solver evaluates today in testing

Line 2 and Line 3 both queue a changeover at 2:15 PM. Four sequencing options scored by projected quality-risk and energy-cost delta:

DATA VIEW
4 No change — first-in, first-served +$410 combined energy cost
3 Stagger changeovers by 20 minutes +$260 combined energy cost
2 Pre-stage Line 3's changeover materials +$90 combined energy cost
1 Shift Line 2's changeover to the next scheduled gap +$15 combined energy cost

WHAT CHANGED

Time returned

~85% of manual monitoring time

Supervisors got the time back for coaching and changeovers, not for screen-watching.

Operating complexity

Reactive breakdown response became a reviewed, prioritized checklist

Where this stands

First system live in 45 days

CHAPTER 03 OF 5

UP NEXT · TEAM VIEW

See what actually changed for the people doing the work, role by role.

Continue to Team View