Platform

Perception at the machine. Reasoning above it. Guardrails around both.

Pulpum has three layers: a GPU mill-edge node that sees the web in real time, an orchestrator that plans and arbitrates between agents, and a policy engine that decides what may actually be written to the DCS. The digital twin sits alongside all three, rehearsing every move first.

Book a mill review Read the docs

3 layers9 agents + orchestratorEdge-first, cloud-optional

run_8f21c4 · agent graph · PM4 10 nodes · 1 approval gate
01 · orchestrator ingest.grade_target ✓ SUCCEEDED 02 · orchestrator twin.simulate ✓ SUCCEEDED 03 · pulp_stock stock.refine ✓ SUCCEEDED 04 · wetend_chem wetend.dose ✓ SUCCEEDED 05 · form_press form.headbox ✓ SUCCEEDED 06 · form_press press.nip ✓ SUCCEEDED 07 · dry_coat dry.steam ✓ SUCCEEDED 08 · defect_inspect inspect.web ✓ SUCCEEDED 09 · orchestrator approve.human ◆ APPROVAL 10 · quality_conf reel.qualify ✓ SUCCEEDED

Live example: Grade change PM4 · 135 gsm kraftliner → 110 gsm testliner, no break, ≤14 min off-spec

Architecture

The loop, end to end

Perceive, plan, simulate, act, verify, log, learn. Every Pulpum run walks the same seven stages regardless of which agent owns it.

  1. Perceive

    Line-scan and IR cameras, QCS scans, DCS tags, lab results and web-monitoring signals are ingested at the mill edge and time-aligned to the reel.

  2. Plan

    The orchestrator turns the declared goal into a typed step graph, assigns each step to an agent, and marks which steps need a human.

  3. Simulate

    Candidate recipes run on the as-run twin — forming, press, dryer, calender — and are ranked on off-spec tonnes, break risk and energy.

  4. Act

    Each write passes the policy engine: tag allow-list, rate and magnitude limits, autonomy level, interlock state, operator presence.

  1. Verify

    The next QCS scan and the vision stream confirm the move landed. If it did not, the step is retried, revised or escalated.

  2. Log

    Request, reasoning, limits applied, human decisions and results are appended to a hash-chained audit log the mill owns.

  3. Learn

    Break events, operator corrections and rejected recommendations become training data for the site-specific model.

Run timeline

A run is the unit of work

Not a dashboard, not an alert — a run. It has a goal, a plan, an owner, a duration, an audit trail and an outcome.

run_8f21c4 PM4 · 6.8 m trim · 1,180 m/min SUCCEEDED
  1. 01 ingest.grade_target SUCCEEDED 0.8 s

    Mill Orchestrator — Pulled the 110 gsm testliner spec, customer tolerances and the standing energy budget from mill MES; locked the target envelope for the run.

  2. 02 twin.simulate SUCCEEDED 34 s

    Mill Orchestrator — Simulated 48 candidate transition recipes on the as-run paper-machine twin — forming, press, dryer and calender — and ranked them on off-spec tonnes, break risk and steam.

  3. 03 stock.refine SUCCEEDED 2 m 10 s

    Pulp-and-Stock — Stepped refiner specific edge load 1.9 → 1.4 Ws/m and pushed freeness toward 412 CSF while consistency held at 3.4%.

  4. 04 wetend.dose SUCCEEDED 1 m 26 s

    Wetend-and-Chemistry — Retention aid trimmed to 214 g/t and sizing to 1.1 kg/t against live charge and turbidity; first-pass retention recovered to 78% inside 90 seconds.

  5. 05 form.headbox SUCCEEDED 1 m 05 s

    Form-and-Press — Re-cut the slice profile across 78 actuators and set jet-to-wire to 0.994 to hold formation index through the basis-weight ramp.

  6. 06 press.nip SUCCEEDED 48 s

    Form-and-Press — Nip load reduced 620 → 540 kN/m to protect the lighter web; post-press dryness landed at 47.1%.

  7. 07 dry.steam SUCCEEDED 3 m 18 s

    Dry-and-Coat — Re-phased the steam schedule across 6 dryer groups and rebalanced the hood; reel moisture converged to 7.4% ±0.19 2σ at 6.1% less steam than the standing recipe.

  8. 08 inspect.web SUCCEEDED continuous

    Defect-and-Inspect — 18 line-scan cameras streaming; two edge-crack precursors detected at the drive side and cleared by a 40 kN/m nip trim before either propagated.

  9. 09 approve.human APPROVAL 52 s

    Mill Orchestrator — Speed ramp 1,180 → 1,245 m/min exceeded the site autonomy threshold. Held for the machine tender; approved by J. Okonkwo at 04:57:12.

  10. 10 reel.qualify SUCCEEDED 1 m 31 s

    Quality-and-Conformance — Reel R-24188 released: full genealogy written (furnish, chemistry, CD profiles, defect map, operator actions) and checked against grade spec.

Tool calls

The agent shows its work

Every tool invocation, argument and result is written to an immutable, human-readable log — and every reasoning step is expandable. Nothing about a run is hidden from the mill.

tool-call stream · run_8f21c4
  1. 04:51:04mes.read_orders({ machine: "PM4", next_grade: "TL-110" })
  2. 04:51:04ok spec locked · basis 110 g/m² · moisture 7.4% ±0.5
  3. 04:51:05twin.simulate({ candidates: 48, horizon: "18 min" })
  4. 04:51:39ok best recipe #31 · est. off-spec 6.4 t · break risk 0.07
  5. 04:51:40policy.evaluate({ recipe: 31, autonomy: "L3" })
  6. 04:51:40ok 9 writes permitted · 1 write requires human approval
  7. 04:51:41dcs.write_setpoint({ tag: "REF4.SEL", from: 1.9, to: 1.4 })
  8. 04:53:51ok freeness 388 → 412 CSF · consistency 3.4%
  9. 04:53:52chem.dose({ retention_aid: "214 g/t", sizing: "1.1 kg/t" })
  10. 04:55:18ok first-pass retention 71% → 78%
  11. 04:55:19dcs.slice_profile({ actuators: 78, jet_to_wire: 0.994 })
  12. 04:56:24ok formation index 84 · CD basis 2σ 1.7 g/m²
  13. 04:56:25vision.stream({ cameras: 18, model: "webdefect-v7" })
  14. 04:56:41precursor: edge crack risk 0.62 · drive side · 1,412 mm
  15. 04:56:42dcs.nip_load({ delta: "-40 kN/m", zone: "press-2" })
  16. 04:56:58ok edge crack risk 0.62 → 0.09 · web stable
  17. 04:56:59approval.request({ action: "speed_ramp", to: "1245 m/min" })
  18. 04:57:12ok approved by J. Okonkwo (machine tender) · logged
  19. 04:58:44mes.log_reel({ reel: "R-24188", genealogy: "full" })
  20. 05:02:46ok run complete · 11m42s · off-spec 6.2 t · breaks 0

Reasoning trace

Plan → thought → action → observation, exactly as the orchestrator recorded it.

GoalMove PM4 from 135 gsm kraftliner to 110 gsm testliner with zer…

Move PM4 from 135 gsm kraftliner to 110 gsm testliner with zero breaks and under 14 minutes of off-spec production.

PlanSimulate the transition on the twin first. Lead with stock and…

Simulate the transition on the twin first. Lead with stock and chemistry so the furnish is right before the basis weight ramps, then trim forming and press, then re-phase drying. Watch the web continuously; hold the speed ramp for a human.

ThoughtThe standing recipe drops basis weight and speed together. On …

The standing recipe drops basis weight and speed together. On the twin that produces a 3-minute moisture excursion because the dryer section unloads faster than the steam schedule can follow. Sequencing the steam re-phase ahead of the basis-weight ramp removes it.

Actiontwin.simulate(candidates=48) → recipe #31 re-phases steam 40 s…

twin.simulate(candidates=48) → recipe #31 re-phases steam 40 s before the basis-weight ramp and holds nip load until post-press dryness confirms.

ObservationRecipe #31 estimates 6.4 t off-spec against 9.8 t for the stan…

Recipe #31 estimates 6.4 t off-spec against 9.8 t for the standing recipe, with break risk 0.07. Two candidates scored lower on steam but raised break risk above the 0.15 site limit and were discarded.

ActionExecute recipe #31 under autonomy level L3: nine setpoint writ…

Execute recipe #31 under autonomy level L3: nine setpoint writes permitted, speed ramp routed to the machine tender.

ObservationEdge-crack precursor at 1,412 mm drive side at 04:56:41. Nip t…

Edge-crack precursor at 1,412 mm drive side at 04:56:41. Nip trim of −40 kN/m in press-2 cleared it in 16 s without a break.

OutcomeRun complete in 11 m 42 s. 6.2 t off-spec, zero breaks, 6.1% l…

Run complete in 11 m 42 s. 6.2 t off-spec, zero breaks, 6.1% less dryer steam than the standing recipe. Reel R-24188 released with full genealogy.

Mill edge

The edge node is the product

Paper moves 20 metres a second. Anything that has to make a decision about the web has to make it locally, deterministically, and in under 100 milliseconds.

What runs at the machine

  • DeepStream + Holoscan pipelines across 8–24 synchronised cameras
  • TensorRT-optimised defect and break-precursor models
  • Triton serving vision, prognostics and grade embeddings side by side
  • Local time-series buffer so a network loss never loses reel history
  • Deterministic rollback by model version, per machine
  • Runs on NVIDIA Jetson Orin today; IGX and on-prem GPU for larger fleets
Digital twin

As-run, not as-designed

A twin built from the machine drawing is a marketing asset. A twin continuously corrected against this week's reels is a control tool.

What the twin simulates

  • Forming and drainage against live furnish and freeness
  • Press dewatering and post-press dryness by nip load
  • Dryer-section heat and mass transfer across all groups
  • Calender nip and coating film split for caliper and gloss
  • Web dynamics and break propagation under draw and tension
  • Grade transitions end to end, 10–100 candidates per change
Accelerated computing

Physical AI, at the machine

Paper moves at 1,200 metres a minute. Perception has to be local, deterministic and fast, so Pulpum runs GPU inference at the mill edge and keeps training, simulation and optimisation in the cloud or on-prem.

Jetson Orin · DeepStream · Holoscan · TensorRT

Edge perception

8–24 synchronised line-scan cameras per machine. Sub-100 ms defect classification, sub-250 ms break-risk updates. [ASPIRATIONAL design targets]

Triton · NIM

Model serving

Vision, time-series prognostics, grade embeddings and process reasoning behind one orchestrator, with deterministic rollback by model version.

DGX / HGX · NeMo

Training

Defect vision, break precursors, wet-end response and drying models trained on reel genealogy, QCS histories and operator corrections.

Omniverse · OVX

Digital twin

GPU-accelerated CFD, drying and web-dynamics simulation of the as-run machine — 10–100 candidate recipes evaluated per grade change.

Omniverse Replicator · Cosmos

Synthetic data

Rare break, contamination, wrinkle and formation-upset variants generated and always validated against real mill events before promotion.

cuOpt

Optimisation

Grade-change sequencing, dryer energy allocation, machine-speed balancing and maintenance windows under production and energy constraints.

Autonomy

Four levels, set per agent and per tag

A mill does not go from manual to unattended in one step. Pulpum makes the level explicit, auditable and reversible at any time.

Autonomy levels and the human role at each
LevelWhat the agent doesWhat the human doesTypical time to reach
L1 · AdvisoryRecommends setpoints and explains whyEnters every change manuallyWeek 1
L2 · SupervisedProposes a write; it executes on approvalApproves each write in the HMIWeek 3–6
L3 · BoundedWrites inside tag, rate and magnitude limitsApproves ramps and grade releasesMonth 2–4
L4 · UnattendedRuns the envelope without promptingSets the envelope; reviews the shift recordMonth 6+ [ASPIRATIONAL]
Multi-agent handoff

Agents negotiate, they do not collide

Two agents will want the same actuator. The orchestrator arbitrates on the run goal, not on who asked first — and the handoff is logged like any other step.

Contested move: press-2 nip load

  1. 01form_press.request(nip −0 kN/m) SUCCEEDED0.2 s

    Form-and-Press wants to hold nip load to protect post-press dryness.

  2. 02defect_inspect.request(nip −40 kN/m) SUCCEEDED0.2 s

    Defect-and-Inspect wants to reduce nip load to clear an edge-crack precursor.

  3. 03orchestrator.arbitrate SUCCEEDED0.4 s

    Break risk 0.62 outranks a 0.4-point dryness loss under the run goal "zero breaks". Defect-and-Inspect wins the actuator for 120 s.

  4. 04form_press.handoff(returned) SUCCEEDED120 s

    Actuator returned; Form-and-Press recovers dryness with vacuum instead. Post-press dryness lands at 47.1%.

Arbitration rules

  • Run goal first. Every request is scored against the declared goal, not the requesting agent's local objective.
  • Safety and breaks outrank quality. A break costs hours; a profile excursion costs minutes.
  • Time-boxed ownership. An agent holds a contested actuator for a bounded window, then must re-justify.
  • Everything is logged. The losing request, the score and the reason all appear in the run record.
  • Deadlocks escalate to a human rather than resolving by timeout.

See the agent roster

Agent roster

Nine agents and an orchestrator

Each agent owns a section of the mill, a bounded tool set, and a measured outcome. They negotiate for shared actuators through the orchestrator — never directly.

agent.pulp_stock

Pulp-and-Stock

Pulping, refining, screening, stock prep

Holds freeness, fiber length and consistency on target across furnish swings — recovered fiber, virgin kraft, or blends — by closing the loop on refiner load, specific edge load and dilution.

dcs.read_refinerlab.freenessdcs.write_setpoint

±4 CSF held on 92% of reels

agent.wetend_chem

Wetend-and-Chemistry

Retention, sizing, charge, additives

Doses retention aid, sizing, starch and defoamer against live charge, turbidity and first-pass retention instead of a fixed recipe — cutting chemistry cost and wet-end upsets.

qcs.read_retentionchem.dosetwin.simulate

11% lower additive spend

agent.form_press

Form-and-Press

Headbox, forming, press section

Trims slice profile, jet-to-wire ratio, vacuum and nip load to hit formation and post-press dryness before the web ever reaches the dryer section.

dcs.slice_profileqcs.read_formationdcs.nip_load

+1.4 pts post-press dryness

agent.dry_coat

Dry-and-Coat

Dryer section, coating, calendering

Schedules steam pressure, hood balance, blade load and calender nip to hit moisture, caliper and gloss at the lowest energy per tonne.

dcs.steam_scheduleqcs.read_moistureenergy.budget

7.9% less dryer steam

agent.defect_inspect

Defect-and-Inspect

Web inspection, break precursors

Fuses line-scan vision and web-monitoring to classify holes, wrinkles, spots, streaks and edge cracks in under 100 ms — and flags break precursors before the sheet goes.

vision.streamwim.replayalarm.raise

68% of breaks predicted >90 s early

agent.profile_quality

Profile-and-Quality

Basis weight, moisture, caliper CD profiles

Runs CD profile control across the full actuator set, keeping 2-sigma inside grade spec through speed changes and grade transitions.

qcs.read_profiledcs.cd_actuatorsspec.check

2σ moisture 0.31 → 0.19

agent.yield_takt

Yield-and-Takt

Yield, broke, energy, grade sequencing

Sequences grade changes, balances machine speed against dryer capacity, and prices every candidate move in tonnes, broke and megajoules.

cuopt.sequencemes.read_ordersenergy.budget

−22% grade-change waste

agent.robot_handling

Robot-and-Handling

Reel, winder and roll handling

Drives reel turn-up, winder set changes, roll wrapping and clamp-truck routing so finishing never starves or blocks the machine.

isaac.plan_pathwinder.set_changewms.route

3.1 h/shift of manual handling removed

agent.quality_conf

Quality-and-Conformance

Right-first-time, genealogy, traceability

Builds the reel genealogy record — furnish, chemistry, profiles, defects, operator actions — and gates release against grade spec and standards.

mes.log_reelspec.checkaudit.write

99.2% release records complete

agent.orchestrator · coordinates all nine

Mill Orchestrator

Planning, arbitration, human approval

Plans the run, arbitrates between agents competing for the same actuator, enforces autonomy level and routes anything above the risk threshold to a human.

plan.composepolicy.evaluateapproval.request

100% of writes policy-checked

Measured

What the platform delivers

Composite pilot results across design-partner machines [ASPIRATIONAL].

<100 ms Defect classification at the mill edge
<250 ms Break-risk update from web signals
48 Twin recipes evaluated per grade change
100% Setpoint writes policy-checked and logged

Figures are design targets and pilot-scoped results [ASPIRATIONAL]. Every number is reproduced from the mill's own reel genealogy, not our telemetry.

Developers

Define an agent, bound it, run it

The Pulpum SDK is typed Python. Tools are declared with schemas and limits; the policy engine enforces them at call time — not in a review meeting.

What you get

  • Typed tool definitions with unit-aware ranges and rate limits
  • Deterministic replay of any historical run against a new model
  • Local twin harness so a recipe is simulated before it is shipped
  • Autonomy policy as code, versioned and reviewed like any other change

Read the docs Developer guide

pm4_dryer_agent.py 
# Bound the dryer agent to six steam groups on PM4.
from pulpum import Agent, Tool, Limit, Autonomy

steam = Tool(
    name="dcs.steam_schedule",
    tags=["PM4.DRY.G1..G6.PRESS_SP"],
    limits=[Limit(max_step="0.15 bar", per="30s")],
)

dryer = Agent(
    id="agent.dry_coat",
    goal="reel moisture 7.4% +/-0.5, min steam",
    tools=[steam, Tool("qcs.read_moisture", read_only=True)],
    # bounded writes; humans still gate ramps
    autonomy=Autonomy.L3,
    # simulate on the twin before every write
    verify="twin",
)

run = dryer.start(machine="PM4", grade="TL-110")
for step in run.stream():
    print(step.name, step.status, step.duration)
CLI

Start a run from anywhere

The same run engine, the same policy checks, the same audit trail — from the terminal, the HMI or the SDK.

pulpum · cli 
$ pulpum run "grade change PM4 to TL-110" --autonomy L3

→ plan composed          10 steps · 1 approval gate
→ twin.simulate          48 candidates · best #31 · risk 0.07
→ policy.evaluate        9 writes permitted · 1 held for human
→ executing              stock.refine ... ok   2m10s
→ executing              wetend.dose .... ok   1m26s
→ executing              form.headbox ... ok   1m05s
→ executing              dry.steam ...... ok   3m18s
! approval required      speed_ramp 1180 → 1245 m/min
→ approved               J. Okonkwo · machine tender · 04:57:12
→ run complete           11m42s · off-spec 6.2 t · breaks 0

$ pulpum runs show run_8f21c4 --format genealogy
From the control room

What machine tenders say

Autonomy earns trust one shift at a time. These are design-partner quotes from pilot deployments [ASPIRATIONAL].

"The first thing that convinced the crew wasn't the control — it was the log. You can scroll back and see exactly why it dropped the nip. Nobody argues with a timestamp."

Machine tender · PM4 · Nordkraft Mills

"We had two people who could do a clean 135-to-110 transition. One retired in March. The twin now does the sequencing and the second one supervises it."

Production manager · Aurora Board

"Break prediction was the wedge. Ninety seconds of warning is the difference between a nip trim and four hours of threading."

Process engineer · Ternvik Paper
Integrations

It speaks mill, not cloud

Pulpum reads and writes through the systems already on the floor. No rip-and-replace, no parallel historian, no new HMI to learn.

QCS

Valmet IQ, ABB 800xA QCS, Honeywell Experion MX

Profiles, scans, lab results

DCS

ABB 800xA, Valmet DNA, Honeywell Experion, Siemens PCS 7

Setpoint reads and guarded writes

Web inspection

WIS/WMS line-scan, IR and transmission cameras

Frames, defect maps, break replays

Mill MES

SAP PP/QM, ABB cpmPlus, custom historians

Orders, grades, reel genealogy

Historian

OSIsoft PI, Aspen IP.21, InfluxDB

Time-series backfill and replay

Robotics

NVIDIA Isaac, winder and wrapper PLCs

Reel, roll and clamp-truck motion

Identity

Azure AD, Okta, on-prem LDAP

SSO, RBAC, named approvers

Edge

NVIDIA Jetson Orin, IGX, on-prem GPU

Sub-100 ms inference at the machine

See all integrations

Guardrails

An agent that can move a nip needs a leash

Pulpum writes to production equipment. Every capability is scoped, every write is policy-checked, and every action is written to an append-only audit log the mill owns.

  • Bounded action space. Each agent can only write to an explicit tag allow-list, inside per-tag rate and magnitude limits.
  • Policy engine before every write. Autonomy level, shift, grade, interlock state and operator presence are all evaluated before a setpoint moves.
  • Human-in-the-loop gates. Anything above the site threshold — speed ramps, grade releases, safety-adjacent moves — waits for a named approver.
  • Immutable audit log. Append-only, hash-chained, exportable, and retained on the mill's own storage.
  • Hard fallback. Loss of the edge node, the network or the model returns control to the DCS's last known-good state within one scan cycle.
  • Tenant and IP isolation. Furnish recipes, grade models and defect libraries never cross a customer boundary. On-prem deployment available.

Compliance posture

StandardScopeStatus
SOC 2 Type IICloud control plane RUNNING In progress [ASPIRATIONAL]
ISO 27001Company-wide ISMS QUEUED Planned [ASPIRATIONAL]
IEC 62443Mill-edge OT security RUNNING Design-aligned
GDPROperator data SUCCEEDED Compliant
ISO 9001 / FSCQuality + chain of custody records SUCCEEDED Supported

Read the security overview

Enterprise

Standardise autonomy across the group

One policy model, one audit trail, one benchmark across every machine in every mill — with the grade and furnish models kept private to each site.

Talk to sales Enterprise details

  • SSO, SCIM and role-based access down to the tag level
  • Group-wide autonomy policy with per-site override and approval chains
  • Cross-mill benchmarking on broke, breaks, steam and first-pass quality
  • VPC, on-prem and fully air-gapped mill-edge deployment options
  • 99.5% control-plane SLA; edge autonomy survives control-plane loss
  • Dedicated deployment engineer and quarterly model review per site
FAQ

Straight answers

The questions mill managers and process engineers actually ask in the first meeting.

Yes, but only within an explicit tag allow-list with per-tag rate and magnitude limits, and only at the autonomy level your site has set. Level 1 is advisory-only: Pulpum recommends and a human enters everything. Most mills spend their first weeks there before enabling supervised writes.

Get started

See the platform on your machine

Start with one paper machine and one measurable baseline. A 90-day mill-edge pilot on break prediction, drying energy or moisture profile shows the number before you commit further.

Book a mill review See pricing