Cut-and-Debone agent
Controls robotic primal cutting, deboning and trimming. Adapts cut path, blade force and seam-following to each animal’s anatomy to maximise saleable yield without bone chips or over-trim.
AI agents
Agents that do the work, not chat about it.
A Slicium agent perceives, decides and acts inside a bounded safety envelope — with grounding, approval gates and a full evidence trail on every action it takes.
Agent roster
Six specialists, one orchestrator.
Grade-and-Yield agent
Senses and predicts grade, marbling, lean/fat ratio, weight and yield from vision and hyperspectral imaging — before and after the cut.
Detect-and-Safety agent
Fuses X-ray, RGB and hyperspectral streams for foreign-material, defect and contamination detection, plus hygiene and pathogen-risk intelligence.
Portion-and-Pack agent
Controls portioning, slicing, weight-giveaway and fixed-weight batching so every pack lands on target instead of above it.
Robot-and-Handling agent
Drives robotic loading, hanging, handling, packing and palletising in cold, wet, washdown environments rated for daily sanitation.
Yield-and-Throughput agent
Optimises line balance, scrap, giveaway and cold-chain flow across the plant, and flags off-spec risk before it reaches the pack.
Agent loop
What happens between two carcasses.
-
1
Perceive
Fuse RGB, depth, hyperspectral, X-ray, scale, line-speed and robot-pose streams into one synchronised view of the carcass and the line.
-
2
Plan
Compute the cut path, blade force and seam, the grade call, the detection verdict and the portion batch — validated in the twin before the blade moves.
-
3
Act
Write back into the robotic cell, portioner, grader, rejector and MES with adaptive, per-carcass control at line speed.
-
4
Verify & learn
Measure realised yield, grade, weight and rejects, log the assurance trail, and feed supervised corrections back into training.
Trust model
Autonomy is granted, never assumed.
Grounded, cited, auditable
Process and food-safety reasoning is retrieved from your cut specifications, product specs, HACCP plans and hygiene procedures — with citations enforced so every answer traces to a source your QA team recognises.
- RAG over specs, standards and procedures
- Citation and grounding checks on every output
- Permission-aware, per-tenant retrieval
- No answer without a source
Human-in-the-loop where it counts
High-impact yield and food-safety decisions route to a named approver. Everything else runs autonomously and reports. The approval graph is part of the audit record, not a separate spreadsheet.
- Role-based approval gates
- Automatic escalation below confidence thresholds
- Override capture as training signal
- Immutable action log
Under the hood
The stack behind the agents.
Edge-first inference
TensorRT-optimised perception on Jetson-class hardware, DeepStream for multi-camera ingest, Holoscan for synchronised X-ray, hyperspectral, scale and robot-pose fusion. Triton serves multiple models per cell with deterministic latency.
Model router
Fine-tuned open models handle high-volume perception; frontier models handle protein-process and food-safety reasoning. The router picks the best model per step and controls COGS.
Vector memory
pgvector for defect, contaminant and cut-recipe retrieval, with a time-series store for cutting and packing telemetry.
Craft memory
Per-plant yield history and per-butcher performance memory capture scarce craft — versioned and scoped per tenant.
Guardrails and evaluation
Golden datasets, LLM-as-judge scoring and twin validation gate every model or prompt change in CI. Schema and safety guardrails run at inference time.
Developer view
Agents you can inspect, test and version.
Every agent step is idempotent, typed and replayable against recorded plant telemetry.
cut_agent.py
# Plan a per-carcass cut, validated in the twin before execution.
from slicium import Agent, Twin, Cell
agent = Agent("cut-and-debone", species="poultry")
async def on_carcass(frame):
anatomy = await agent.perceive(frame) # bone, seam, fat/lean
plan = await agent.plan(anatomy, spec="front-half-v7")
sim = await Twin.validate(plan) # predicted yield + risk
if sim.confidence < 0.94 or sim.bone_chip_risk > 0.01:
return agent.escalate(reason="below autonomy gate")
return await Cell.execute(plan, audit=True)slicium-cli — plant edge
$ slicium agents status --line 4
cut-and-debone autonomous conf 0.971 1,412/hr
grade-and-yield autonomous conf 0.964
detect-and-safety autonomous conf 0.992 0 escapes
portion-and-pack advisory conf 0.921
$ slicium twin validate --spec front-half-v7
simulated 12,000 carcasses · predicted yield 82.4% (±0.3)
gate: PASS — promoting to supervised autonomyThe same interfaces power the console, the CLI and the CI pipeline.
Autonomy levels
What each level permits.
| Level | Agent behaviour | Human role |
|---|---|---|
| Shadow | Predicts every decision, acts on none | Reviews accuracy against the current baseline |
| Advisory | Recommends in real time | Accepts or rejects; corrections become training data |
| Supervised | Acts inside a bounded envelope | Approves high-impact and food-safety decisions |
| Graduated | Acts autonomously per qualified workflow | Supervises exceptions and owns the autonomy gate |
Reliability
Agent performance targets.
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
Supervision
See what every agent is doing, and why.
Saleable yield82.4%+4.1% vs baseline
Giveaway0.7%−61% this week
FM rejects120 escapes
Line uptime99.4%+2.2 pts
Line 4 · Front-half deboning1,412 /hr83.1% yieldAutonomous
Line 2 · Breast portioning2,980 /hr0.6% giveawayAutonomous
Line 7 · X-ray screening3,140 /hr4 rejectsAdvisory
Line 1 · Primal fabrication640 /hr78.9% yieldShadow
Illustrative console data from a pilot deboning line. [PLACEHOLDER]
Agent governance
An agent is only as good as its evidence.
SOC 2 Type II (in progress)HACCP-alignedUSDA-FSISEU 853/2004GDPRISO 27001 (planned)
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.
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.
Adoption
How teams learned to trust them.
“The line does not care that every bird is different — Slicium does. We stopped programming a machine and started supervising an operator that adapts to each carcass.”
“Giveaway was the quiet leak nobody could close. Watching the portioner hold target within a couple of grams, shift after shift, changed the economics of the whole pack line.”
“What sold my team was the audit trail. Every reject links back to the frame, the model version and who approved the autonomy level. That is what a recall investigation actually needs.”
Design-partner quotes are composite and pending publication approval. [PLACEHOLDER]
Agent FAQ
How autonomy behaves in practice.
That variability is the whole point. Fixed automation fails because it repeats one motion; Slicium perceives each animal’s anatomy — bone position, seam location, fat/lean boundary — and plans a cut path, blade force and seam-follow specific to that carcass. The cut is validated against a digital twin that predicts yield before the blade moves.
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]
Start with one line. Prove the yield.
Run a paid pilot on a single processing line with a defined yield, giveaway or labour success metric. Shadow mode first, autonomy only when the numbers earn it.