Home / Approach & technology
Approach & technologyA modular verification system — not a black box.
StunAssure combines video, environmental and process sensors, stunner parameters, and expert-reviewed indicators into a conservative, explainable risk-detection and audit platform. The goal is not to replace stunning technology — it is to make humane stunning measurable, adaptable, and practical across economic contexts.
Four practical layers, fused into one signal.
The innovation is not one AI camera — it is the layered architecture that combines weak signals and stays fail-safe.
Protocol layer
Manual checklist, stun-to-kill timer, species profile, and standard operating procedures — usable offline and in low-resource settings.
Sensor layer
Temperature, conductivity/salinity, dissolved oxygen where needed, stunner settings, and batch timing — logged without modifying existing stunners.
Decision-support layer
Conservative, explainable risk scoring that returns a simple, fail-safe signal: likely acceptable, uncertain — check manually, or intervention needed.
Camera-assist layer
Detects visible warning signs — coordinated movement, continued ventilation, recovery indicators — as an optional higher tier, never as a consciousness verdict.
Honesty first. The system does not claim that video alone proves consciousness. It is designed to flag risk and support validated decision-making with domain experts — never to certify "safe" on weak evidence. See our Responsible AI boundary →
The verification stack, by readiness
The science largely exists — only the verification layer is new. Backbone is ready now; Echo-Stun is the funded research stretch.
Inputs → risk engine → operator alert → audit report.
A buildable architecture: practical inputs feed a conservative risk engine that produces a clear operator signal and a saved batch record.
From observation to welfare assurance.
A multi-layer assurance workflow for detecting, documenting, and reducing fish-welfare risks during stunning and slaughter. The system does not claim to determine consciousness — it surfaces observable welfare-risk indicators and the cases that need human review.
- Observe
- Measure
- Synchronize
- Detect risk
- Explain
- Review
- Improve
-
01
Observe
Cameras capture visible post-stun movement, handling, posture, and process flow.
Why Surfaces observable signs that may warrant review.
-
02
Measure
Sensors and process logs record water temperature, conductivity, pH, timing, batch data, and stunner settings where available.
Why Stunning effectiveness depends on real conditions, not equipment alone.
-
03
Synchronize
Video, sensor data, operator notes, and process events are aligned on one timeline.
Why Creates structured evidence instead of isolated observations.
-
04
Detect risk
Rules and optional models flag observable welfare-risk indicators — unexpected movement, delayed handling, process deviations, or missing data.
Why Supports early detection of problems, not consciousness verdicts.
-
05
Explain
Each alert shows the reason: what was detected, when it happened, and which data supported the flag.
Why Keeps the system transparent and reviewable.
-
06
Review
Operators or welfare advisors validate flagged clips and uncertain cases.
Why Keeps expert human judgment central to every decision.
-
07
Improve
Reports help facilities correct problems, refine procedures, and document welfare assurance.
Why Turns monitoring into continuous welfare improvement.
Affordable entry points — not lower welfare standards.
Tiering creates affordable pathways toward better verification and process control. Lower-cost tiers are more conservative; they should never falsely certify fish as insensible.
Lite
Small farms & low-resource settings
- Checklist app
- Stun→kill timer
- Species SOPs
- Manual logs
Sensor Box
Medium farms & processors
- Water sensors
- Stunner logging
- Drift alarms
- Audit reports
Camera Assist
Advanced sites
- Edge camera
- Visual warning detection
- Higher automation
Validation Mode
Researchers & certifiers
- Expert labels
- Validation protocol
- Benchmark data
Illustrative target ranges, not fixed prices.
Six signal layers, one fused risk picture.
No single signal is trusted on its own. The risk engine fuses weak, complementary signals — and the parameters it logs are exactly the ones the welfare science says determine whether a stun actually works. See the underlying evidence →
| Layer | Signals captured | Why it matters |
|---|---|---|
| Visual | Post-stun movement, posture, equilibrium, ventilation / mouth movement, eye & body response (where visible) | Non-invasive monitoring — but visible signs alone never prove unconsciousness. |
| Stunner / process | Voltage, current, frequency, exposure time, waveform, throughput, batch density | Electrical-stun efficacy depends directly on these parameters. |
| Environmental | Water temperature, dissolved O₂, salinity, pH, turbidity, conductivity | Conductivity sets the field strength a stun needs; the rest give context. |
| Expert / welfare | Advisor-reviewed indicators, species-specific rules, risk thresholds | Species-specific welfare is central to EFSA- and WOAH-style assessment. |
| Audit | Logs, video clips, sensor readings, alerts, generated reports | Evidence for farms, processors, vets, certifiers and regulators. |
| Learning | Model improvement from expert-reviewed cases | Builds a versioned dataset and improves over time — human in the loop. |
Built from low-cost, proven, open components.
Nothing here is exotic. The whole system runs on commodity edge hardware and mature open-source software, so it stays affordable and can operate fully offline at a remote farm or vessel.
| Layer | Components | Role |
|---|---|---|
| Edge compute | Raspberry Pi 5 → NVIDIA Jetson Orin Nano | On-site inference; offline-capable; Pi for the Lite tier, Jetson when vision is added. |
| Sensors & MCU | ESP32 · DS18B20 temp · EC/salinity · pH · dissolved-O₂ · turbidity · current/voltage | Low-cost environmental and process logging without modifying the stunner. |
| Vision | USB / IP industrial camera · OpenCV · YOLOv8/v11 | Optional edge detection & tracking of visible post-stun warning signs. |
| Data & messaging | MQTT · TimescaleDB / InfluxDB · PostgreSQL | Time-synchronised, multimodal evidence on one timeline. |
| Application | FastAPI · Next.js / React · Grafana · Docker | Operator dashboard, alerts, batch audit reports, containerised deployment. |
| ML ops | PyTorch · scikit-learn · Label Studio/CVAT · MLflow · DVC · ONNX/TensorRT | Conservative models, human-in-the-loop labelling, versioned datasets. |
Buy cheap, build smart. We build the data pipeline and dashboard first on simulated stunner data and low-cost sensors; expensive hardware — industrial cameras, and EEG/VER rigs used only for scientific validation — enters at later tiers. General slaughterhouse welfare-camera systems already exist (e.g. AI4Animals, Argus); StunAssure differs by being fish-specific, low-cost, explainable, multimodal and expert-reviewed.