Notes

Notes from the loop, the granulator, the bagging line and the control room.

Short essays on fertilizer manufacturing and autonomy, written from the thinking behind the product. No author bylines yet: the team publishes names when people join publicly.

Six notesFertilizer manufacturing and autonomyNo sales register

run_scn_01 · agent graph · scenario 10 nodes · 1 approval gate
01 · orchestrator order.receive ✓ SUCCEEDED 02 · orchestrator twin.simulate ✓ SUCCEEDED 03 · reform_synth loop.control ✓ SUCCEEDED 04 · react_evap evaporate.control ✓ SUCCEEDED 05 · granulate granulate.control ✓ SUCCEEDED 06 · coat_condition condition.cool ✓ SUCCEEDED 07 · defect_inspect inspect.granules ✓ SUCCEEDED 08 · orchestrator approve.human ◆ APPROVAL 09 · blend_bag blend.bag ✓ SUCCEEDED 10 · quality_conf release.genealogy ✓ SUCCEEDED

Illustrative scenario: Complex run · one urea granulation campaign from synthesis gas to bagged lot · ammonia loop, urea synthesis and evaporation, granulator, dryer and cooler, NPK bagging line. Not a measured run.

Notes

In the order we wrote them

Origin

The plant drifts all day, and nobody can see it in time

A fertilizer complex drifts all day. The ammonia loop, the evaporators and the granulators run on fixed setpoints and operator craft, and the cost of a bad shift shows up as off-spec tonnes, dust, recycle and energy nobody gets back. Quality is checked by periodic lab assay and a hand lens, hours after the granule was made. By then the off-spec tonnes are already in the pile.

The tools the plant has measure and log. APC packages tune a setpoint, analyzers assay a sample, the MES tracks what happened. None of them close the loop across the complex to adaptive conversion, granulation and blend control. The gap was never sensors. It was action.

That is the opening Nitreum was designed for: connect perception to action, start on one unit with one number, and earn each step of autonomy against the plant's own records.

Workforce

The craft is retiring faster than plants can replace it

The process engineers who can hold a complex on spec are retiring, and the plants they leave behind are remote, giant and energy-hungry. Conversion gets checked against a shift report, granule size against a sieve, moisture against a lab result. When a caking problem reaches a customer, it becomes a claim.

Our answer is not to replace the engineer. It is to make their corrections the thing the plant learns from: every approval, every override and every note in assist mode becomes a labeled event in a per-site memory that is versioned, tenant-scoped and never shared across customers.

Granulex is the clearest case. Per-product and per-granulator memory encodes what the engineer knows about this product on this line, so the next campaign starts from what the last one taught.

Simulation

Why a twin has to be as-produced, not as-designed

Producers commit to a grade change, a recipe or a changeover sequence and only discover a conversion miss, a wet granule bed, an emissions excursion or a takt bottleneck after tonnes, energy and CO2 are spent. An as-designed model of the equipment does not help, because the plant never runs as designed: the catalyst has aged, the feedstock changed, the ambient temperature is not what the curve assumed.

Nitrotwin starts from the live plant state and the coupled physics of the ammonia loop, evaporation, granulation, blending and emissions, and simulates the campaign as it will actually be produced. The result is a plan Synthex, Granulex and Blendex can execute, and a record of why the discarded candidates were discarded.

The same machinery generates rare dust, caking, wet-granule and upset scenarios that are too rare and costly to collect, always validated against real plant events before a model is allowed to learn from them.

Inspection

Vision that feeds the granulator

Standalone inspection is a report. It tells you the distribution drifted or the dust rose and leaves the granulator doing what caused it. So plants run extra recycle and extra coating, because an out-of-spec lot is a return and extra recycle is only energy.

Granvis is built on the opposite premise: a fault is a signal about the granulator, the dryer or the coating state, and the signal is worth more than the report. Size and shape, dust, caking precursors and broken prills are detected at the plant edge, and the drift goes back to Granulate-and-Prill to trim the spray or the recycle inside the approved window.

It is also why Granvis is usually the first wedge: continuous quality visibility is immediate, measurable on the plant's own assays and QC records, and visible within weeks of shadow mode.

Synthesis

Energy is the product's biggest cost, and the loop runs on fixed setpoints

Ammonia synthesis is one of the most energy-intensive processes on earth, and energy is the dominant cost of nitrogen fertilizer. Most loops run on fixed setpoints with a correction after the shift report, because a loop upset is a safety event and a few percent of energy is only money.

Synthex holds conversion from live loop telemetry and a conversion model, inside an envelope the process-safety review approved, and lets a twin-validated plan take energy per tonne below the fixed curve. Purge and trip settings are read and verified, never written, and any move beyond the envelope waits for a named approver.

The commercial case follows the physics: energy per tonne and emissions per tonne are numbers every nitrogen plant already tracks, which makes outcome pricing on verified reduction straightforward.

Principles

Trust before autonomy

Autonomy in a fertilizer complex is earned in stages. Shadow mode predicts and writes nothing. Assist mode recommends and a person decides. Bounded control writes low-risk granulation, conditioning and blending setpoints inside documented limits. Safety-instrumented systems, interlocks, permit limits and release decisions never move to the agent.

Every output is grounded and cited to the spec, the safety procedure or the emissions permit it relied on. Every action is written to an immutable, assurance-grade audit log. Model and prompt changes are gated by golden datasets and model-as-judge evaluations in CI before they reach a plant.

These are the guiding principles in our roadmap, and they are the reason the first release is one wedge workflow with a review console and an audit log, not a complex-wide program.

Run timeline

The scenario most notes refer to

Every Nitreum run is an ordered, inspectable sequence. This is run_scn_01, an illustrative scenario: Complex run · one urea granulation campaign from synthesis gas to bagged lot · ammonia loop, urea synthesis and evaporation, granulator, dryer and cooler, NPK bagging line. Scenario complete · lot inside spec · one human approval on the ammonia loop.

run_scn_01 scenario · not measured SUCCEEDED
  1. 01 order.receive SUCCEEDED 0.8 s

    Complex Orchestrator · Pulled the grade spec, the nutrient target, the granule spec for size distribution, hardness and moisture, the emissions permit limits, the safety-instrumented envelope and the standing changeover window from the plant MES and the historian; locked the target envelope for the campaign.

  2. 02 twin.simulate SUCCEEDED 41 s

    Complex Orchestrator · Simulated 24 candidate loop, evaporation, granulation and sequencing plans on the Nitrotwin as-produced complex twin; ranked them on conversion, energy per tonne, predicted granule distribution, recycle ratio, emissions margin and takt, and discarded every candidate that thinned the safety or permit margin.

  3. 03 loop.control SUCCEEDED continuous

    Reform-and-Synthesize · Held ammonia-loop conversion and reformer temperature inside the approved envelope from live loop telemetry and the conversion model, trimming setpoints toward the lowest energy per tonne the plan allowed; purge and trip settings were read, never written.

  4. 04 evaporate.control SUCCEEDED continuous

    React-and-Evaporate · Held urea melt concentration at the granulation target from the online concentration analyzer and the evaporation model, moving evaporator setpoints inside the recipe window rather than waiting for the next lab assay.

  5. 05 granulate.control SUCCEEDED continuous

    Granulate-and-Prill · Tuned recycle ratio, spray and bed temperature from the live Granvis size distribution and moisture signal, each move validated on the twin first; the granule distribution stayed inside the spec band while recycle trended down.

  6. 06 condition.cool SUCCEEDED continuous

    Coat-and-Condition · Held dryer and cooler outlet temperature and moisture inside the storage window and set the anti-caking coating rate to the predicted caking risk for the product and the season; nothing moved outside the conditioning envelope.

  7. 07 inspect.granules SUCCEEDED continuous

    Defect-and-Inspect · 12 camera streams at the plant edge: granule size and shape, dust, caking precursors and broken prills on the granulator discharge, cooler and bagging belts; two dust spikes were attributed to one spray nozzle and Granulate-and-Prill corrected it inside its window.

  8. 08 approve.human APPROVAL 2 m 14 s

    Complex Orchestrator · The lowest-energy candidate wanted a purge-rate change on the ammonia loop beyond the bounded envelope. Held for the shift process engineer and the process-safety approver, who approved a smaller step with the safety-instrumented envelope untouched and a tighter review interval.

  9. 09 blend.bag SUCCEEDED continuous

    Blend-and-Bag · Held the NPK blend ratio to the nutrient spec from Granvis nutrient sensing, checked segregation on the belt and held bag weight to target; three overweight bags were attributed to one dosing gate and the gate was trimmed inside the approved window.

  10. 10 release.genealogy SUCCEEDED 1 m 10 s

    Quality-and-Conformance · Lot released: conversion and energy per tonne, melt concentration, granule distribution, hardness and moisture, nutrient assay, emissions readings, inspection events, bag weights and every human decision written to the lot genealogy and checked against spec and permit.

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 plant. The log below is the scenario run, not a measured one.

tool-call stream · run_scn_01
  1. 05:10:02mes.read_order({ campaign: "U-4471", grade: "UREA-46", pack: "BAG-50" })
  2. 05:10:02ok grade spec locked · permit limits v4 · SIS envelope read-only
  3. 05:10:03historian.read_baseline({ unit: "AMM-1", window: "30d" })
  4. 05:10:03ok energy per tonne baseline loaded · catalyst age noted
  5. 05:10:04twin.simulate({ candidates: 24, horizon: "campaign" })
  6. 05:10:45ok best plan #9 · permit margin held · lowest energy per tonne
  7. 05:10:46policy.evaluate({ plan: 9, stage: "bounded" })
  8. 05:10:46ok 7 writes permitted · 1 write requires human approval
  9. 05:10:47dcs.write_setpoint({ tag: "AMM1.LOOP_T", basis: "conversion+energy" })
  10. 05:14:30ok conversion on target · energy per tonne trending down
  11. 05:20:00analyzer.read_concentration({ unit: "EVAP-2" })
  12. 05:20:00ok melt concentration inside window · evaporator setpoint held
  13. 05:31:12granvis.read_distribution({ unit: "GRAN-1", metric: "size_moisture" })
  14. 05:31:12ok distribution inside spec band · recycle ratio trimmed after twin check
  15. 05:31:13vision.stream({ cameras: 12, models: ["size_shape", "dust", "caking", "bag_defect"] })
  16. 06:02:48dust spike · cooler belt · attributed to spray nozzle 3
  17. 06:02:49defect.attribute({ event: "dust_spike", cause: "granulator.nozzle_3" })
  18. 06:02:49ok granulate notified · spray trimmed inside window · second spike, then clean
  19. 06:02:50dcs.write_setpoint({ tag: "COAT1.RATE", basis: "moisture+caking_risk" })
  20. 06:02:50ok coating rate set · moisture inside storage window
  21. 06:02:52approval.request({ action: "loop.purge", reason: "beyond bounded envelope" })
  22. 06:05:06ok approved by the shift process engineer and the process-safety approver · smaller step · SIS untouched
  23. 07:40:10plc.write_setpoint({ tag: "BLEND1.RATIO", basis: "nutrient+segregation" })
  24. 07:40:10ok nutrient ratio on spec · bag weight on target · gate 2 trimmed
  25. 09:18:30mes.release_lot({ campaign: "U-4471", genealogy: "full" })
  26. 09:19:40ok lot released · scenario complete

Reasoning trace

Goal, plan, thought, action, observation, exactly as the orchestrator would record it.

GoalRun campaign U-4471 from synthesis gas to bagged lot: ammonia …

Run campaign U-4471 from synthesis gas to bagged lot: ammonia loop at target conversion per unit of energy, melt at exact concentration, granules on spec with minimum dust and recycle, conditioned for storage, bagged to weight and released with full genealogy and emissions readings.

PlanSimulate the whole campaign on the twin first. Lead with conve…

Simulate the whole campaign on the twin first. Lead with conversion so evaporation gets the right melt and the granulator gets the right concentration. Verify emissions and the safety envelope continuously. Let the granulator chase the size band only inside the approved window; hold any ammonia-loop move beyond the envelope for a human.

ThoughtThe standing practice runs the loop and the granulator on fixe…

The standing practice runs the loop and the granulator on fixed setpoints and corrects after the assay. With live conversion, concentration and granule distribution the loop and the granulator can sit at target. The saving is energy, recycle and dust, not safety margin: the safety-instrumented envelope and the permit limits are never candidates for optimization.

Actiontwin.simulate(candidates=24) → plan #9: conversion on target, …

twin.simulate(candidates=24) → plan #9: conversion on target, melt concentration on spec, granule distribution inside the band, permit margin unchanged, lowest predicted energy per tonne.

ObservationPlan #9 wants a purge-rate change on the ammonia loop beyond t…

Plan #9 wants a purge-rate change on the ammonia loop beyond the bounded envelope. The policy engine routes that single move to the shift process engineer and the process-safety approver and permits the other seven writes.

ActionExecute plan #9 at the bounded stage: loop temperature, evapor…

Execute plan #9 at the bounded stage: loop temperature, evaporator, granulator, coating and blend writes proceed; emissions are verified continuously; the purge move waits.

ObservationApproval granted at 06:05:06 with a smaller step and the safet…

Approval granted at 06:05:06 with a smaller step and the safety-instrumented envelope untouched. Two dust spikes were attributed to spray nozzle 3 and corrected by Granulate-and-Prill inside the approved window.

OutcomeScenario complete. The lot is released with conversion and ene…

Scenario complete. The lot is released with conversion and energy per tonne, melt concentration, granule distribution, nutrient assay, emissions readings and the human decision in the genealogy. No measured plant result is claimed; this is the run the platform is designed to execute.

Topics

What we write about

Synthesis

Conversion, energy per tonne, the ammonia loop, evaporation and safety-gated control.

Granulation

Size distribution, moisture, hardness, dust, recycle, coating and craft capture.

Quality and emissions

Granule vision, nutrient analyzers, CO2, NOx and ammonia slip, and the twin.

Trust

Process safety, audit logs, autonomy stages and what never leaves a person.

Autonomy

Four stages, set per loop

A plant does not go from manual to bounded control in one step. Nitreum makes the stage explicit per loop, auditable and reversible at any time, and some decisions never leave the person. On the ammonia side every stage is safety-gated.

Autonomy stages and the human role at each
StageWhat the agent doesWhat the person doesWhen
ShadowPredicts conversion, concentration, granule quality, nutrient content, energy, emissions and off-spec risk; writes nothingRuns the plant as today; compares predictions to assay, historian and engineer outcomesFirst phase with every design partner
AssistRecommends bounded loop, evaporation, granulation, coating and blend moves with the evidenceA process engineer approves or rejects each oneAfter accuracy is proven against the baseline
BoundedWrites low-risk granulation, conditioning and blending setpoints inside documented recipe, quality, safety, emissions, equipment and rollback limitsApproves anything above the limits and every ammonia-loop move beyond the envelope; reviews the shift recordAfter the safety and ROI gates agreed up front
ExceptionsEscalates anything outside its envelope, safety-adjacent moves and release decisionsAlways decides these; the agent never doesPermanent
Design targets

Figures we cite most

The numbers Nitreum is engineered against, from the GPU utilization plan.

50–120 ms p95 edge inference for granule size, shape, dust, caking and bag-defect frames
8–40 cameras per synthesis, granulation or bagging unit, 1080p or 4K at 15 to 30 FPS
1–10 Hz sensor fusion of vision, analyzer, historian, emissions, acoustic and thermal streams
2–5 model families fine-tuned per design-partner cohort
500–2,000 GPU-hours of fine-tuning per design-partner cohort
50k–250k synthetic rare-upset images and scenarios per cohort, labeled as synthetic

Design targets from the GPU utilization plan, all marked aspirational in the architecture. Nothing on this page is a measured plant result; measurements start when the first design-partner unit runs in shadow mode.

Guardrails

An agent that can move a setpoint needs a leash

Nitreum writes to production equipment in a plant that handles ammonia. Every capability is scoped, every write is policy-checked, every output is cited, and every action is written to an append-only audit log the plant owns.

  • Bounded action space. Each agent can only write to an explicit tag allow-list, through approved DCS or APC interfaces, inside per-tag rate and magnitude limits, at the autonomy stage the plant has set.
  • Safety systems are read, never written. Safety-instrumented systems, interlocks, trip settings and emissions permit limits stay with the plant's process-safety program and the people who own it.
  • Policy engine before every write. Autonomy stage, safety envelope, interlock state, grade, shift and operator presence are evaluated before a setpoint moves. Any ammonia-loop move beyond the envelope waits for a named approver.
  • Grounded and cited. Every recommendation carries a citation to the spec, procedure or permit it relied on. Retrieval is tenant-isolated and permission-aware.
  • Immutable audit log. Append-only, exportable and retained on the plant's own storage, aligned to the plant's quality, emissions and process-safety record-keeping.
  • Hard fallback. If the edge node, the network or a model stops, the plant keeps running on its existing DCS, APC and safety systems.
  • Tenant and IP isolation. Recipes, process settings and plant IP never cross a customer boundary. On-prem and plant-edge deployment available.

Assurance posture

Assurance and certification status
ItemScopeStatus
SOC 2 Type I readinessCloud control plane QUEUED Roadmap, first 6 months
SOC 2 Type IICloud control plane QUEUED Roadmap, 6 to 12 months
Process-safety and emissions proceduresGrounding and record-keeping RUNNING Design basis for retrieval and the audit log
Per-tenant isolationData, memory, models RUNNING Design rule
On-prem and plant-edge deploymentIP- and safety-sensitive sites RUNNING Design option

Nitreum supports your process-safety, quality and emissions records. It is not a certification and it does not replace your safety program or management of change.

Read the security overview

Integrations

It speaks plant, not cloud

Nitreum reads and writes through the systems already in the control room. No parallel historian, no new HMI, no rip-and-replace. Connectors are built against each design partner's stack first; long-tail integrations are out of scope for the first release.

DCS, APC and historian

Reformer, ammonia loop, reactor and evaporator control and the plant historian

Setpoint reads, guarded writes, safety systems untouched

Granulation and prilling control

Granulator, prill tower, dryer, cooler and coating drum controls

Recycle, spray, temperature and coating setpoints inside windows

Online analyzers

Nutrient, moisture and concentration analyzers, NIR and granule analyzers

Continuous quality state, fused at the edge

Emissions sensing

CO2, NOx and ammonia-slip sensors and permit records

Read and verified; excursions predicted, limits never optimized against

Blending and bagging

NPK blender, dosing gates, bagging and loading lines

Ratio and weight setpoints inside approved windows

Cameras and robots

Fixed cameras on granule, belt and bagging stations; Isaac-driven palletizing and loading cells

Frames, defect events, bounded handling

Plant MES, LIMS and quality

Orders, grade specs, lab assays, lot genealogy

Targets in, release records out

Identity (planned)

SSO and role-based access; SOC 2 on the roadmap

Named approvers, scoped roles

See all integrations

FAQ

Straight answers

The questions complex operations managers, process engineers and HSE leads ask in the first meeting.

Only at the autonomy stage your plant has set, and only to an explicit allow-list of low-risk loops inside documented limits, through the approved DCS or APC interfaces. Every design partner starts in shadow mode, where Nitreum predicts and writes nothing. Assist mode follows, where a process engineer approves each recommendation. Bounded autonomy on granulation, conditioning and blending loops comes after the safety and ROI gates agreed up front.

Get started

Get new notes

Ask to be added to the list. Notes only, no drip campaign.