Official implementation of the Best Paper Award-winning framework at PRO-VE 2025.
The Collaborative Intelligence Component is a web-based implementation of an Agentic Framework for Industry 5.0. It demonstrates a novel approach to Human-in-the-Loop (HITL) decision-making at the Edge, integrating Large Language Models (LLMs) to automate data curation and enable seamless, one-click model recalibration.
Keywords: Edge AI · Industry 5.0 · Human-AI Collaboration · Human-in-the-Loop · Agentic Frameworks · Collaborative Intelligence · Large Language Models · Explainable AI · MQTT · Smart Manufacturing
- Citation
- Archived Software DOI (Zenodo)
- Video Demonstration
- Scientific Contribution & Capabilities
- System Architecture
- Usage Guide
- Customization
- Implementation Notes & Scope
- Troubleshooting
- Related Links
- Contributing
- Acknowledgment
- License
If you use this code or framework in your research, please cite the peer-reviewed paper. GitHub also exposes a "Cite this repository" button (top-right of the repo page), powered by the included CITATION.cff.
Plain text:
Martinez-Gil, J., Pichler, M., Bountouni, N., Koussouris, S., Márquez Barreiro, M., & Gusmeroli, S. (2026). An Agentic Framework for Rapid Deployment of Edge AI Solutions in Industry 5.0. In L. M. Camarinha-Matos, A. Ortiz, X. Boucher, & A. Lucas Soares (Eds.), Hybrid Human-AI Collaborative Networks (PRO-VE 2025), IFIP AICT vol. 771 (pp. 55–68). Springer, Cham. https://doi.org/10.1007/978-3-032-05681-8_4
BibTeX (published version, please cite this one):
@inproceedings{martinezgil2026agentic,
author = {Martinez-Gil, Jorge and Pichler, Mario and Bountouni, Nefeli and
Koussouris, Sotiris and Barreiro, Marielena M{\'a}rquez and Gusmeroli, Sergio},
title = {An Agentic Framework for Rapid Deployment of Edge {AI} Solutions in Industry 5.0},
booktitle = {Hybrid Human-AI Collaborative Networks (PRO-VE 2025)},
editor = {Camarinha-Matos, Luis M. and Ortiz, Angel and Boucher, Xavier and Lucas Soares, Antonio},
series = {IFIP Advances in Information and Communication Technology},
volume = {771},
pages = {55--68},
year = {2026},
publisher = {Springer, Cham},
doi = {10.1007/978-3-032-05681-8_4},
isbn = {978-3-032-05681-8}
}Read the paper: Springer (published) · arXiv (open access)
Watch the system in action, featuring real-time MQTT data processing and ChatGPT 4o integration:
https://www.youtube.com/watch?v=AR8F8U-QXhM
This repository implements the "Collaborative Intelligence" layer described in the associated paper. It bridges the gap between static Edge AI models and dynamic industrial environments through:
Traditional "black box" deployments give operators no way to review an automated call before it reaches the line. This tool instead provides a Dynamic Validation Interface where human experts can review, correct, and curate AI predictions in real time.
The system integrates Generative AI (ChatGPT 4o) to act as an intelligent agent. It not only labels data but provides Explainable AI (XAI) reasoning for why a prediction was flagged as "OK" or "Non-OK," reducing the cognitive load on human operators.
The component features a feedback loop by design: validated data (from humans or the GenAI agent) is meant to recalibrate the Edge AI model with a single click, letting the system adapt to data drift without offline retraining cycles. In this reference build the click is wired to a placeholder confirmation; see Implementation Notes & Scope for what that means in practice.
To address the data scarcity common in industrial settings, the system leverages LLMs to generate additional training samples based on expert-curated positive and negative examples.
The solution follows a modular architecture designed for Edge deployment:
- Frontend: Lightweight HTML/CSS/JS interface for low-latency interaction.
- Communication Layer: MQTT-based telemetry (compatible with HiveMQ, Mosquitto) for real-time sensor streams.
- Intelligence Layer:
- Edge Model: Handles immediate inference.
- Cloud Agent: Connects to GenAI APIs (Chatbase/OpenAI) for higher-order reasoning.
- Visualization: Plotly.js for dynamic, real-time performance monitoring.
Message flow in the demo: a batch of readings goes out over MQTT, comes back as a prediction, and a human or the GenAI agent can validate it before an optional recalibration step.
flowchart LR
CSV["data.csv<br/>(batch upload)"] -->|Send CSV| Client
Edge["External edge device<br/>(optional)"] -->|publish: inputTopic+'-'| Broker(("MQTT broker"))
Client["client.html<br/>Collaborative Intelligence UI"] -->|publish: inputTopic| Broker
Broker -->|inputTopic| Server["server.html<br/>demo edge processor"]
Server -->|publish: outputTopic| Broker
Broker -->|outputTopic, inputTopic+'-'| Client
Client -->|GenAI request| AI["Chatbase API"]
AI -->|explanation text| Client
Client -->|review & correct| Expert(("Domain expert"))
Expert -->|validated label| Client
Client -->|Recalibrate Model| Model["Edge AI model"]
- Modern Web Browser (Chrome, Firefox, Edge).
- Access to an MQTT broker over WebSockets (e.g. the public HiveMQ broker used in the default config).
- Outbound internet access:
client.htmlandserver.htmlload MQTT.js, Plotly.js, and Google Fonts from a CDN, so both files need a network connection even when opened locally.
| File | Purpose |
|---|---|
app/client.html |
Main Collaborative Intelligence interface (monitor, validate, recalibrate). |
app/server.html |
Companion data publisher that streams sample readings to the broker to simulate an edge device. |
app/config.json |
MQTT broker URL, topics, and input/output feature mapping. |
data.csv |
Example industrial batch dataset. |
The repository root also carries citation and archival metadata (CITATION.cff, codemeta.json, LICENSE) and the award image (best-paper.jpg). app/data.csv and app/data.csv0 are additional local CSV samples used during development; they are not required to run the demo. logo.png is duplicated at the root for project listings and previews, while the running app loads app/logo.png.
- Launch: Open
app/client.htmlin your browser. - (Optional) Simulate a device: Open
app/server.htmlin a second tab to publish sample telemetry to the broker. - Configure: Adjust
app/config.json(or load it from the interface) to point at your MQTT broker and define your input/output features. - Connect: Connect to the broker to start the real-time MQTT stream.
- Collaborate:
- Monitor incoming predictions in the Dynamic Table.
- Use the "GenAI" action to auto-generate reasoning for anomalies.
- Manually correct targets if necessary.
- Improve: Click "Recalibrate Model" to trigger the recalibration hook (see Implementation Notes & Scope).
Define your MQTT topics and feature vectors:
{
"brokerURL": "wss://broker.hivemq.com:8884/mqtt",
"inputTopic": "input",
"outputTopic": "output",
"inputs": [
{"name": "Vibration_Sensor_X", "value": ""},
{"name": "Temperature_C", "value": ""},
{"name": "Pressure_PSI", "value": ""}
],
"outputs": [
{"name": "Target", "value": ""}
]
}- Styling: Fully customizable CSS for white-labeling.
- Model Backend: The JavaScript logic is modular;
predict()functions can be swapped for TensorFlow.js or ONNX runtimes.
This repository is the reference implementation that accompanies the PRO-VE 2025 paper. A few things are worth knowing before pointing it at a real production line:
- Recalibration is a hook, not a trainer. The Recalibrate Model button in
app/client.htmlcurrently shows a confirmation message. It marks where a retraining or redeployment call belongs, so it can be wired to a real pipeline. - GenAI needs its own credentials. The GenAI button calls the Chatbase REST API through
sendMessageToChatabase(). TheapiKeyandchatbotIdshipped in the file are placeholders, so replace them with real values before the button returns real explanations. - The default broker is public.
app/config.jsonpoints at the publicbroker.hivemq.cominstance so the demo works with no setup. That broker is unauthenticated and visible to anyone subscribed to the same topics, so swap in a private, authenticated broker before sending real sensor or product data. - One link is machine-specific. The "Input Analysis" button in
app/client.htmlopenshttp://localhost:8501/?file_url=C:/Users/martinez/Documents/GitHub/ci-component/data.csv, a local path used during development. Point it at your own analysis tool, or remove it, before sharing a deployed copy.
- "Connect to Broker" stays disabled. The button only activates after a valid
config.jsonis loaded through Load Config; on success, the app calls the broker connection automatically. - Status never leaves "Not connected to broker." Check the browser console for a WebSocket error. Public brokers like HiveMQ can refuse connections under load; retry, or point
brokerURLat a broker under your own control. - GenAI returns "Error fetching AI response." This is expected until real Chatbase credentials are set (see Implementation Notes & Scope).
- Recalibrate Model only shows an alert. That is the current behavior of the reference build; see Implementation Notes & Scope.
- The page loads with no charts and no MQTT activity.
client.htmlandserver.htmlfetch their libraries from a CDN, so confirm the browser has an active internet connection even for a local file.
- Published paper (Springer): https://doi.org/10.1007/978-3-032-05681-8_4
- Open-access preprint (arXiv): https://arxiv.org/abs/2510.25813
- Project reference implementation: https://github.com/AI-REDGIO-5-0/ci-component
- AI REDGIO 5.0 project: https://www.airedgio5-0.eu
Issues and pull requests are welcome, particularly around the recalibration hook, broker security defaults, and additional edge-model backends. Fork the repository, create a feature branch, and open a pull request that describes the change and how it was tested. For questions about the research behind the component, see Citation.
This work is supported by the AI REDGIO 5.0 project: "Regions and (E)DIHs alliance for AI-at-the-Edge adoption by European Industry 5.0 Manufacturing SMEs" under the European Union's Horizon Europe research and innovation programme, Grant Agreement No. 101092069.
Released under the MIT License. © 2023–2026 Jorge Martinez-Gil.


