Every model, prompt and policy that can touch a unit carries a version, and the run record names the versions that produced it.
Every release that can touch a unit, written down.
Nothing has shipped to a plant yet. This is the release line the first design partners will see, in the order it is planned, plus the policy that governs how any release reaches a unit.
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.
The planned release line
Status is honest: everything is queued. Order follows the first-release plan.
-
v0.1Connectors and shadow mode QUEUED
Connectors to the wedge unit's DCS or APC, historian, online analyzers, emissions sensors, cameras, lab system and plant MES. Read-only. Predictions logged against assay, historian and engineer outcomes; nothing written.
-
v0.2Review console and audit log QUEUED
The console where process engineers and HSE leads approve, correct and annotate; the immutable, assurance-grade quality, emissions and safety audit log from day one.
-
v0.3Assist mode QUEUED
Recommendations for synthesis, evaporation, granulation, conditioning or blend moves with the evidence and citation behind each one; approvals and corrections become labeled training events.
-
v0.4Bounded autonomy QUEUED
Guarded writes to low-risk granulation, conditioning and blending loops inside documented recipe, quality, safety, emissions, equipment and rollback limits, after the safety and ROI gates.
-
v0.5Nitrotwin: the plan before the run QUEUED
A surrogate twin of one unit, the ammonia loop or a granulator, calibrated against as-produced lots; pre-run validation of one agent's moves and one changeover; a validated synthetic rare-upset set.
-
v0.6Second unit in the complex QUEUED
An adjacent agent in the same complex with deeper integrations; SSO and role-based access; SOC 2 Type II in progress; emissions and safety assurance packages.
How a release reaches a unit
Golden datasets and model-as-judge evaluations gate model and prompt changes in CI.
Nothing reaches a plant until its predictions have been compared to that plant's assay, historian and engineer outcomes in shadow mode.
Rollback by model version; a Nitreum outage returns the plant to its existing control.
The reference scenario each release is tested against
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.
-
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.
-
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.
-
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.
-
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.
-
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.
-
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.
-
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.
-
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.
-
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
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.
What the tool log will show per release
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.
- 05:10:02mes.read_order({ campaign: "U-4471", grade: "UREA-46", pack: "BAG-50" })
- 05:10:02ok grade spec locked · permit limits v4 · SIS envelope read-only
- 05:10:03historian.read_baseline({ unit: "AMM-1", window: "30d" })
- 05:10:03ok energy per tonne baseline loaded · catalyst age noted
- 05:10:04twin.simulate({ candidates: 24, horizon: "campaign" })
- 05:10:45ok best plan #9 · permit margin held · lowest energy per tonne
- 05:10:46policy.evaluate({ plan: 9, stage: "bounded" })
- 05:10:46ok 7 writes permitted · 1 write requires human approval
- 05:10:47dcs.write_setpoint({ tag: "AMM1.LOOP_T", basis: "conversion+energy" })
- 05:14:30ok conversion on target · energy per tonne trending down
- 05:20:00analyzer.read_concentration({ unit: "EVAP-2" })
- 05:20:00ok melt concentration inside window · evaporator setpoint held
- 05:31:12granvis.read_distribution({ unit: "GRAN-1", metric: "size_moisture" })
- 05:31:12ok distribution inside spec band · recycle ratio trimmed after twin check
- 05:31:13vision.stream({ cameras: 12, models: ["size_shape", "dust", "caking", "bag_defect"] })
- 06:02:48dust spike · cooler belt · attributed to spray nozzle 3
- 06:02:49defect.attribute({ event: "dust_spike", cause: "granulator.nozzle_3" })
- 06:02:49ok granulate notified · spray trimmed inside window · second spike, then clean
- 06:02:50dcs.write_setpoint({ tag: "COAT1.RATE", basis: "moisture+caking_risk" })
- 06:02:50ok coating rate set · moisture inside storage window
- 06:02:52approval.request({ action: "loop.purge", reason: "beyond bounded envelope" })
- 06:05:06ok approved by the shift process engineer and the process-safety approver · smaller step · SIS untouched
- 07:40:10plc.write_setpoint({ tag: "BLEND1.RATIO", basis: "nutrient+segregation" })
- 07:40:10ok nutrient ratio on spec · bag weight on target · gate 2 trimmed
- 09:18:30mes.release_lot({ campaign: "U-4471", genealogy: "full" })
- 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.
What comes after the first release
From the roadmap, by horizon. Every item is a plan, not a result.
| Horizon | Planned |
|---|---|
| 0–6 months | Two to four design partners on the wedge workflow; MVP: connectors, agent loop, review console, assurance-grade quality, emissions and safety audit log; accuracy and energy, off-spec and emissions ROI proven in shadow then assist; security and process-safety baseline with SOC 2 Type I readiness; IP-safe plant-edge deployment |
| 6–12 months | Design partners convert to paid; graduated autonomy on low-risk synthesis and granulation loops; Nitrotwin hit-spec-before-run; second unit per complex; SOC 2 Type II; SSO and RBAC; first producer logos; emissions and safety assurance packages |
| 1–3 years | The full synthesis, reaction and evaporation, granulation and blending loop; single complex to multi-site producer to global group; process-licensor and control-system-vendor channels; international expansion; benchmark data products |
| 3–10 years | The autonomous system of action for fertilizer manufacturing, with adjacent scopes from reforming through bagging and loading, green-ammonia integration and a platform and API ecosystem |
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.
| Stage | What the agent does | What the person does | When |
|---|---|---|---|
| Shadow | Predicts conversion, concentration, granule quality, nutrient content, energy, emissions and off-spec risk; writes nothing | Runs the plant as today; compares predictions to assay, historian and engineer outcomes | First phase with every design partner |
| Assist | Recommends bounded loop, evaporation, granulation, coating and blend moves with the evidence | A process engineer approves or rejects each one | After accuracy is proven against the baseline |
| Bounded | Writes low-risk granulation, conditioning and blending setpoints inside documented recipe, quality, safety, emissions, equipment and rollback limits | Approves anything above the limits and every ammonia-loop move beyond the envelope; reviews the shift record | After the safety and ROI gates agreed up front |
| Exceptions | Escalates anything outside its envelope, safety-adjacent moves and release decisions | Always decides these; the agent never does | Permanent |
Physical AI, at the plant edge
A fertilizer complex is remote, continuous and often restricted from cloud links, so perception has to be local and fast. Nitreum runs GPU inference at the plant edge and keeps training, simulation and optimization in the cloud, in a private cloud or on-prem. Items marked planned are on the roadmap, not running.
Edge perception
8 to 40 camera feeds per unit through DeepStream pipelines and TensorRT detectors and segmenters for granule size and shape, dust, caking, broken prills and bag defects; Holoscan fuses vision with analyzer, historian, emissions, acoustic and thermal streams at 1 to 10 Hz. Design target: 50 to 120 ms p95 inference.
Model serving
Vision, time-series, RAG and process-reasoning models behind one orchestrator, with canary deployment and per-site routing. NVIDIA AI Enterprise is the planned runtime for regulated deployments.
Training and reasoning (planned)
Granule vision, time-series prognostics and NeMo process-reasoning models fine-tuned from labeled images, process histories, engineer corrections and event logs. Design target: 2 to 5 model families and 500 to 2,000 GPU-hours per cohort.
Digital twin (planned)
The Nitrotwin complex twin: physics-informed surrogates for the ammonia loop, evaporation, granulation, blending and emissions, with GPU-accelerated process and CFD sweeps run before a plant change.
Synthetic data (planned)
Rare dust, caking, wet-granule, broken-prill, lighting, belt-contamination, emissions-spike and upset scenarios generated and labeled as synthetic, validated against real plant events before a model learns from them.
Optimization
Throughput, recycle balancing, energy and emissions allocation and grade-changeover sequencing under constraints, recomputed on the GPU when the complex is disrupted.
Data engineering
ETL over historian, quality, emissions and vision data for feature generation, quality correlation and cross-site benchmarks on the GPU.
Robotic handling
Bag palletizing, bulk loading, materials handling and inspection rovers, with perception and motion policies built and validated in simulation before they move a pallet.
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
| Item | Scope | Status |
|---|---|---|
| SOC 2 Type I readiness | Cloud control plane | QUEUED Roadmap, first 6 months |
| SOC 2 Type II | Cloud control plane | QUEUED Roadmap, 6 to 12 months |
| Process-safety and emissions procedures | Grounding and record-keeping | RUNNING Design basis for retrieval and the audit log |
| Per-tenant isolation | Data, memory, models | RUNNING Design rule |
| On-prem and plant-edge deployment | IP- 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.
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.
Reformer, ammonia loop, reactor and evaporator control and the plant historian
Setpoint reads, guarded writes, safety systems untouched
Granulator, prill tower, dryer, cooler and coating drum controls
Recycle, spray, temperature and coating setpoints inside windows
Nutrient, moisture and concentration analyzers, NIR and granule analyzers
Continuous quality state, fused at the edge
CO2, NOx and ammonia-slip sensors and permit records
Read and verified; excursions predicted, limits never optimized against
NPK blender, dosing gates, bagging and loading lines
Ratio and weight setpoints inside approved windows
Fixed cameras on granule, belt and bagging stations; Isaac-driven palletizing and loading cells
Frames, defect events, bounded handling
Orders, grade specs, lab assays, lot genealogy
Targets in, release records out
SSO and role-based access; SOC 2 on the roadmap
Named approvers, scoped roles
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.
Safety-instrumented systems, interlocks, trip settings and permit limits are not optimization variables. Nitreum reads and verifies them, writes them to an immutable audit log and never writes to them. Any ammonia-loop move outside the bounded envelope waits for a named process engineer and a process-safety approver, and management of change stays with your plant. Nitreum is grounded in your process-safety and emissions procedures; it is not a certification and does not replace them.
Nitreum sits above your existing automation as a supervisory layer. If it stops, the plant keeps running on its current DCS, APC and safety systems. A Nitreum outage degrades the plant to the way it runs today, not to a stop.
The first phase is shadow mode on one unit: Nitreum reads your historian, analyzers, emissions sensors, lab assays and MES while you run the plant as usual, and we measure its predictions against your assay, historian and engineer outcomes. Recommendations start when accuracy is proven against that baseline, and improve as site-specific history and engineer corrections accumulate.
Recipes, process settings, product and emissions models and plant IP are tenant-scoped and never used to train another customer's models. Retrieval is multi-tenant isolated and permission-aware. Cross-site learning, where a customer opts in, is federated and IP-isolated. On-prem and plant-edge deployment is available for sensitive sites.
You are, the same as with any control strategy. That is why every write is policy-checked, bounded, logged and reversible, every output carries a citation to the spec or procedure it relied on, and anything above your threshold waits for a named approver. The audit log records the request, the reasoning, the limits applied and the human decision.
As a design partner, on one unit, with one hard metric: energy per tonne on the ammonia loop, granule size distribution and moisture on a granulator, off-spec and recycle tonnes, or emissions variance. We instrument the unit in shadow mode, measure the baseline, move to assist mode when accuracy is proven, and expand only when the number holds.
Full multi-unit autonomy across the entire complex, long-tail integrations beyond the design partners' stacks, and self-serve onboarding. The first release is one wedge workflow, its connectors, the agent loop with human checkpoints, the review console and the assurance-grade quality, emissions and safety audit log.
Follow the release line
Design partners see each release before it reaches their unit. Ask to be one.