Quickstart
Install the edge runtime, register a line, connect your first system and run in shadow mode.
Documentation
Everything needed to run the loop.
Guides, references and runbooks for engineers, integrators and quality teams deploying Slicium on live production lines.
Browse
Documentation sections.
Plant-edge deployment
Cluster sizing, GPU allocation, enclosure and network guidance for washdown environments.
Agent API
Perceive, plan, validate, execute, escalate — with types, errors and replay semantics.
Connector SDK
OPC UA, EtherNet/IP, Modbus, REST and vendor SDK adapters, plus certification steps.
Twin & validation
Simulation setup, synthetic rare-event generation and CI gating configuration.
Security & compliance
Identity, isolation, audit-trail schema and evidence export for HACCP and FSIS audits.
Core concepts
Slicium models a processing line as a sequence of cells. Each cell exposes sensors, actuators and a safety envelope. An agent is bound to one workflow across one or more cells, and the plant orchestrator sequences agents, holds the envelope and writes the assurance record.
The control loop
Every carcass passes through the same four phases: perceive, plan, act and verify. The loop is closed — realised outcomes feed back as supervision, which is what makes the models improve with every animal processed.
loop.txt
frame → perceive → plan → twin.validate → gate → execute → verify → logAutonomy gates
An agent cannot raise its own autonomy level. Promotion from shadow to advisory, supervised and graduated autonomy requires measured accuracy above the configured gate, a passing twin validation, and an explicit human approval recorded in the assurance graph.
Deployment model
The default deployment runs entirely on the plant edge. A K3s cluster hosts the runtime, Triton serves models per cell, and stop paths remain hardwired and independent of software. Cloud connectivity is optional and used for training, fleet management and reporting.
Latency budgets
Budgets are configured per cell and monitored continuously. A missed budget degrades the cell to a safe state rather than delivering a late action. Typical targets are sub-50 ms for reject decisions and sub-100 ms for cut-path updates. [ASPIRATIONAL]
Evidence & audit
Every action writes an immutable record linking the sensor frame, model version, confidence, autonomy level, approver, actuation and realised outcome. Records are exportable in formats accepted by HACCP and USDA-FSIS reviews.
Examples
Copy, paste, run.
quickstart.sh
# Install the edge runtime and register a line.
curl -sSL https://get.slicium.com | sh
slicium login --plant northfold
slicium line register --id line-4 --species poultry \
--workflow deboning
# Connect equipment, then start in shadow mode.
slicium connect opcua://10.4.2.11:4840 --role portioner
slicium run --mode shadow --line line-4slicium-cli — plant edge
$ slicium audit query --lot 4471
carcass 4471-0912 cut front-half-v7 model cut-poultry@3.4.1
confidence 0.973 autonomy supervised approver m.ellison
yield 83.2% fm_rejects 0 frames s3://plant/edge/4471-0912
$ slicium audit export --lot 4471 --format fsis
wrote fsis-4471.pdf (18 pages, 214 linked frames)
Guides
The four guides most teams read first.
-
1
Install the runtime
Provision edge nodes, allocate GPUs, and validate enclosure and network placement with your OT team.
-
2
Connect equipment
Register cutting, grading, detection, portioning and MES systems, and confirm read and write capabilities.
-
3
Baseline in shadow
Run predictions without actions until accuracy against your baseline is statistically meaningful.
-
4
Promote through gates
Advance to advisory, supervised and graduated autonomy with recorded approvals at each step.
Reference
Agent API at a glance.
| Call | Purpose | Returns |
|---|---|---|
| perceive(frame) | Fuse sensors into a structured view of the carcass and line state | Anatomy, grade, detections, geometry |
| plan(view, spec) | Produce a cut, grade, detection or portion plan against a spec | Plan with confidence bounds |
| twin.validate(plan) | Simulate the plan and predict yield, spec compliance and risk | Predicted yield, risk, gate verdict |
| execute(plan) | Write back into the cell within the safety envelope | Actuation record |
| escalate(reason) | Route the decision to a human and degrade to a safe state | Escalation record |
Security docs
Controls documented for auditors.
SOC 2 Type II (in progress)HACCP-alignedUSDA-FSISEU 853/2004GDPRISO 27001 (planned)
Per-tenant isolation
Every plant gets isolated data, models and vector stores. Cut recipes and species models never cross a tenant boundary — federated learning shares patterns, never your product IP.
Assurance-grade audit trail
Immutable logs link sensor frames, model version, autonomy level, approvals, overrides and outcomes for every agent action — built for recall defence and model governance.
Fail-safe by design
Blade, robot and line stops are deterministic and independent of the cloud. Graceful degradation returns the cell to a safe state if perception confidence drops.
Connector reference
Supported systems and protocols.
Cutting & deboning
- Robotic primal cutting cells
- Deboning and trimming lines
- Blade force and seam controllers
- Cell safety and E-stop interlocks
Grading & inspection
- Vision and hyperspectral graders
- Carcass grading cameras
- X-ray and metal detectors
- Checkweighers and rejectors
Portioning & packing
- Portioners and slicers
- Fixed-weight batching
- Packing and palletising robots
- Labelling and traceability
Operational targets
What the docs hold us to.
4.1%
Average saleable-yield uplift on pilot deboning lines [PLACEHOLDER]
61%
Reduction in weight giveaway per fixed-weight pack
38 ms
Median foreign-material reject decision at the plant edge
99.9%
Plant-edge runtime uptime target with fail-safe line stop
Runbooks
Operational runbooks.
Sanitation cycle handling
Safe shutdown, enclosure protection and restart validation around daily caustic washdown.
Model drift response
Detecting supplier, seasonal and breed drift, and the correct escalation path.
Foreign-material escalation
What happens in the first ten minutes after a suspected escape, and what evidence to pull.
Docs FAQ
Common implementation questions.
Slicium runs a graduated autonomy model: shadow, then advisory, then supervised autonomy, each gated by measured accuracy and twin validation. Below the confidence threshold the cell degrades to a safe state, escalates to a human, or routes the product to rework — it never guesses on a food-safety decision.
No. The perception and control loop runs entirely on the plant edge. Cloud is used for training, fleet management and reporting, and can be disabled for on-prem or air-gapped deployments. Federated learning shares model improvements without exposing plant recipes, supplier data or tenant images.
A typical pilot runs one line for 8–12 weeks: two weeks of integration and shadow-mode baselining, four to six weeks of advisory operation, then graduated autonomy against the agreed success metric. Connectors for common cutting, grading, X-ray, portioning and MES systems are pre-built. [PLACEHOLDER]
Cells are specified for cold, wet, IP69K washdown environments with daily caustic sanitation. Perception enclosures, cabling and edge compute are selected for condensation, temperature swing and high-pressure cleaning, and validated with your sanitation crew during commissioning.
Need a hand with the first line?
Our solutions engineers commission alongside your team — on site, in your washdown gear.