Pricing
Request a pilot Read the docs
Developers

Physical AI you can version, test and roll back.

Slicium is an engineering platform first: typed agent interfaces, replayable telemetry, twin validation in CI, and signed artefacts delivered to the plant edge.

Read the docsTalk to an engineer
simulated cut sequence · predicted yield 82.4%
Developer experience

Built like infrastructure, not a demo.

Typed agent APIs

Every agent step is typed, idempotent and replayable against recorded plant telemetry. No hidden state, no untestable prompts.

Plant-edge runtime

A K3s edge cluster with Triton multi-model serving, TensorRT-optimised perception and deterministic latency budgets.

Connector SDK

Write connectors for equipment we have not met yet: OPC UA, EtherNet/IP, Modbus, REST and vendor SDKs.

Twin in CI

Every model or program change is simulated against thousands of synthetic carcasses before it can be promoted.

Replay and time travel

Reproduce any decision from the recorded frame, model version and configuration. Debug the past, not a summary of it.

Signed OTA delivery

Artefacts are signed, versioned and staged, with instant rollback under your own change-control process.

Quick start

From connector to first prediction.

Point the runtime at a line, register your equipment, and start in shadow mode.

connector.py 
# Register a portioner and stream weights into the loop.
from slicium.connect import Connector, OPCUA

portioner = Connector(
    name="line-2-portioner",
    transport=OPCUA("opc.tcp://10.4.2.11:4840"),
    capabilities=["read:weight", "write:portion_target"],
)

@portioner.on("pack")
async def on_pack(evt):
    # Shadow mode: predict, log, do not act.
    pred = await portioner.agent.predict(evt.geometry)
    portioner.log(predicted=pred.grams, actual=evt.grams)
slicium-cli — plant edge 
$ slicium init --plant northfold --line 2
✓ edge runtime detected (k3s, 2 nodes, 4 GPUs)
✓ connectors: portioner, checkweigher, MES

$ slicium run --mode shadow
streaming 2,980 packs/hr · mae 3.1 g · drift ok

$ slicium gate check portion-and-pack
accuracy 0.962 (gate 0.950) · twin PASS · eligible: advisory
Architecture

Where the compute lives.

Edge: perception and control

Jetson-class inference inside washdown-rated enclosures runs the loop with deterministic latency. DeepStream handles multi-camera ingest; Holoscan synchronises X-ray, hyperspectral, scale, line speed and robot pose.

  • TensorRT-optimised perception at 30–120 FPS [ASPIRATIONAL]
  • Triton multi-model serving per cell
  • Independent hardwired stop paths
  • Offline-capable — no cloud in the control loop
IP69K washdown cell

Cloud: training, twin and fleet

Training, synthetic data generation, evaluation and fleet management run centrally. Models ship down; only telemetry and approved corrections come up — or nothing at all in air-gapped deployments.

  • Omniverse Replicator synthetic rare events
  • cuOpt optimisation for cut sequence and batching
  • Golden-dataset and LLM-as-judge evaluation gates
  • Federated learning without tenant data movement
core
Stack

What Slicium is built on.

LayerTechnology
Edge runtimeRust and C++ runtime on K3s, NVIDIA IGX and Jetson
PerceptionPyTorch, TensorRT, DeepStream, Metropolis, Holoscan
SimulationOmniverse, Replicator, Cosmos, cuOpt
ServingTriton, NIM, NeMo with Retriever and Guardrails
BackendGo, Python FastAPI, NestJS control plane
DataPostgreSQL with pgvector, TimescaleDB, Redis
AgentsLangGraph orchestration, model-agnostic router
OpsGitHub Actions, Terraform, ArgoCD, edge OTA, OpenTelemetry
Delivery

How a change reaches the blade.

  1. 1

    Commit

    Model, program or connector change lands with tests and a recorded-telemetry replay suite.

  2. 2

    Evaluate

    Golden datasets, judge scoring and drift checks run in CI. Regressions block the build.

  3. 3

    Simulate

    The twin runs the change against thousands of synthetic carcasses including rare anatomy and contamination.

  4. 4

    Stage & ship

    Signed artefacts roll to a canary cell, then the line, then the cohort — with instant rollback.

For engineers

The things that actually matter at 3am.

Full tracing, end to end

OpenTelemetry traces span sensor frame, model inference, decision, actuation and outcome. Langfuse captures reasoning steps. When a line stops, you know why in seconds — not after pulling four vendors onto a call.

Graceful degradation

Every failure mode has a defined safe state. Perception loss, network loss and model-server loss all degrade predictably.

Runbooks included

Commissioning, sanitation, drift and escalation runbooks ship with the deployment.

Named engineers

Enterprise agreements include named solutions engineers who know your lines.

No silent model swaps

Model versions are pinned per line. Nothing changes on your floor without your change-control approval.

Connectors

Protocols and systems supported.

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

Plant systems

  • MES and production monitoring
  • OPC UA / Modbus / EtherNet-IP
  • Historians and time-series stores
  • ERP and order specs

Quality & food safety

  • HACCP control points
  • QMS and CAPA workflows
  • Traceability and lot genealogy
  • Sanitation and hygiene records

Identity & workflow

  • SSO / SAML / SCIM
  • Role-based approval gates
  • Ticketing and on-call alerting
  • Shift handover surfaces

See the integration reference

Targets

Engineering budgets we hold ourselves 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
Security

Secure by construction.

SOC 2 Type II (in progress)HACCP-alignedUSDA-FSISEU 853/2004GDPRISO 27001 (planned)

Data residency & on-prem

Run fully on-premise on the plant edge, in your VPC, or hybrid. Sensitive producers can keep every frame inside the facility.

SSO, SAML and RBAC

Enterprise identity, role-based approval gates, and separation of duties between operations, quality and engineering.

Signed, versioned edge OTA

Models and runtimes ship as signed artefacts with staged rollout, instant rollback and twin-validated gating in CI.

Read our security overview

Start here

Documentation entry points.

Reference

Agent API reference

Perceive, plan, validate, execute and escalate — the five calls every agent implements.

Docs

Guide

Writing a connector

Map an unfamiliar portioner, grader or MES into the loop with the connector SDK.

Docs

Build on the plant edge.

Get the developer preview, the connector SDK and a walkthrough of the edge runtime with one of our engineers.

Talk to an engineerRead the docs