WORLD MODEL
PROCESS WORLD MODEL
A multimodal model over inline sensors, formula structure and batch tags that predicts where viscosity, particle size, pH proxy and stability land before the batch finishes.
PLATFORM
One runtime.
Every vessel.
Every line.
All five Serumon agents share one plant-edge autonomy runtime, one multimodal perception and process-world-model stack, and one multi-tenant formula, shade and telemetry platform. Deploy on a line, expand to the plant, standardise across sites.
1–10 HzCONTROL CADENCE
50–150msP95 DECISION
K3sON-SITE ORCHESTRATION
OTAMODEL DELIVERY
PLANT NETWORK ├─ IGX / JETSON ORIN vessel cluster A (emulson) ├─ JETSON ORIN spectro + vision (shadera) ├─ JETSON + IGX filler / robot cell (fillix) ├─ EDGE K3s agent runtime, OPC-UA bridge └─ uplink (optional) training + fleet ops OFFLINE BEHAVIOUR CONTINUES CONTROL MODEL ROLLOUT SIGNED, STAGED, REVERSIBLE DATA EGRESS POLICY-GATED PER TENANT
OPC-UAMODBUSMES / BATCHSPECTROPHOTOMETRYCHECKWEIGHERSHOMOGENISERSROBOT CELLSLIMSHISTORIANSERP
ARCHITECTURE
The stack, layer by layer
LAYER 01SENSE
Connectors to vessels, homogenisers, spectrophotometers, fillers, inspection and MES over OPC-UA and APIs.
LAYER 02MODEL
Multimodal perception plus a learned process world model per formula and vessel.
LAYER 03DECIDE
Constrained agents: MPC and RL policies bounded by validated envelopes.
LAYER 04ACT
Signed write-back into setpoints, gates, rejects and robot programs.
LAYER 05PROVE
Immutable GMP-grade audit, evidence packs and the review console.
RUNTIME
What runs inside the plant
DETERMINISTIC CONTROL
Inference and control loops are pinned to the edge so a WAN outage never stalls a vessel. Target p95 decision latency is 50–150 ms.
- TensorRT-quantised models
- Pre-allocated GPU memory
- Watchdog + safe-state fallback
BOUNDED ENVELOPES
Every actuatable parameter carries a min, max and rate limit signed off by process and quality owners before autonomy is enabled.
- Per formula and per vessel
- Versioned and diffable
- Change requires dual approval
HUMAN CHECKPOINTS
The review console is where engineers approve, correct and annotate. Corrections are training data, not tickets.
- Approve / reject / amend
- Reason codes captured
- Feeds the next model refresh
IMMUTABLE AUDIT
Perception, decision, action and approver are appended to a tamper-evident log aligned to ISO 22716 expectations.
- Signed events
- Full replay of any batch
- Export to your QMS
OFFLINE SURVIVAL
The edge keeps a local model cache, local state store and local queue. Reconnection reconciles rather than replays blindly.
- Local Timescale buffer
- Idempotent sync
- No cloud in the control path
FLEET OPERATIONS
Signed OTA rollouts, staged canaries per vessel and single-click rollback across a multi-site estate.
- Canary by vessel class
- Drift + health telemetry
- One-click revert
DEPLOYMENT
From survey to bounded autonomy
-
01
SITE SURVEY
Map vessels, homogenisers, spectros, fillers, robots, MES and network segmentation. Identify the wedge workflow with the clearest ROI.
-
02
INSTRUMENT
Install edge GPU nodes, wire connectors, verify tag quality and start recording. Nothing is written back in this phase.
-
03
BASELINE + SHADOW
Measure current shade holds, giveaway, changeover and yield. Run agents in shadow against your chemists and lab results.
-
04
ASSIST → BOUNDED
Enable recommend mode, then approve-to-act, then bounded autonomy per formula once accuracy and envelopes are validated.
MODEL STACK
What is actually learning
PERCEPTION
SPECTRAL + VISION
Fine-tuned spectral regression for ΔE against the brand standard, plus high-speed defect and fill-level vision across formats.
CONTROL
MPC + RL POLICIES
Model-predictive and reinforcement-learned control over shear, temperature, vacuum, time and filler setpoints, always clamped by the envelope.
PHYSICS
PINN SURROGATES
Physics-informed emulsification and rheology surrogates give the twin useful fidelity at interactive speed instead of overnight CFD.
RISK
MICRO + STABILITY
Time-series prognostics that fuse process signatures with lab assays to score micro and shelf-life risk early in the batch life.
LANGUAGE
GROUNDED REASONING
Retrieval over SOPs, batch records and specs with enforced citations for deviation write-ups and formulation Q&A. No ungrounded answers.
NVIDIA STACK
GPU is not optional here
| AGENT | EDGE | DATA CENTRE | SDKs |
|---|---|---|---|
| EMULSON | Jetson Orin / IGX | DGX / HGX | CUDA, TensorRT, Modulus, NeMo, Triton |
| SHADERA | Jetson Orin / Thor (IGX) | DGX | Metropolis, Holoscan, TensorRT, Replicator |
| FILLIX | Jetson + IGX Orin | DGX | Isaac, TensorRT, Triton |
| STABION | Jetson | DGX / HGX | RAPIDS, NeMo, Triton, NIM |
| TWYNEX | IGX | DGX + OVX | Omniverse, Modulus, cuOpt, Cosmos |
WHY EDGE GPU
CPU-only cannot serve this loop
A single vessel streams inline sensor channels at 1–10 Hz while a spectral head and multiple cameras run at line speed. On top of that the agent must run a physics-informed surrogate and an MPC solve inside the same control window.
That is concurrent multimodal inference plus simulation plus optimisation, every cycle, with a hard deadline. It is a GPU workload at the edge and a DGX-class workload in training.
- Concurrent multimodal inference per vessel and line.
- Physics-informed emulsion simulation inside the control window.
- Real-time state estimation and MPC / RL policy evaluation.
- High-speed defect vision at full line rate without dropped frames.
- Large-scale training over spectra, telemetry, formulas and batch records.
LATENCY BUDGET / CONTROL CYCLE
- SENSOR FUSION
- 8 ms
- PERCEPTION
- 22 ms
- STATE ESTIMATE
- 14 ms
- SURROGATE SIM
- 38 ms
- MPC SOLVE
- 26 ms
- WRITE-BACK
- 9 ms
- P95 TOTAL
- <150 ms
DATA PLATFORM
Multi-tenant by construction
ISOLATION
Per-tenant data and model scoping, with an on-prem option for manufacturers who will never let formula IP leave the site.
TIME SERIES
TimescaleDB for vessel and line telemetry, retained at full resolution for the batches that matter.
RETRIEVAL
PostgreSQL + pgvector for shade standards, defect images, SOPs and batch records with permission-aware filtering.
EVENTS
An event-driven agent runtime with idempotent steps, replay and human-approval checkpoints as first-class citizens.
OPERATING TARGETS
What the runtime is engineered to hold
99.9%EDGE UPTIME TARGET
<150msP95 CONTROL DECISION
10 HzMAX SENSOR CADENCE
0CLOUD CALLS IN CONTROL PATH
ENGINEERING TARGETS FOR DESIGN-PARTNER DEPLOYMENTS [ASPIRATIONAL UNTIL MEASURED IN PRODUCTION].
PLATFORM FAQ
Architecture questions
Yes. The control path never depends on the cloud. In an air-gapped deployment, models are delivered as signed artefacts and telemetry stays on site; training then happens on customer-provided GPU capacity or on exported, contractually scoped datasets.
Through OPC-UA, Modbus, REST and file-drop connectors, plus direct historian reads. Serumon is designed to complement batch and MES systems: they remain the system of record, Serumon becomes the system of action.
The node fails to a safe state and control reverts to your existing DCS or PLC recipe. Serumon is additive: losing it degrades you to the status quo, never to an unsafe state.
Every model version is immutable, tied to a validation record, and rolled out through staged canaries. Model changes follow your change-control process, and the audit log lets you replay any batch against the exact model version that ran it.
ENGINEERING DEEP DIVE
Bring your OT diagram
The fastest technical qualification is a whiteboard session with your process, quality and OT leads. We map vessels, lines, tags and network segmentation, and tell you honestly which wedge is worth instrumenting first.