GSoC 2026 — Interest in Project #34: Agentic AI for Predictive Maintenance of Cognitive Cities Critical Infrastructure

[johyeongseob]

Hello @hmoustaf, @gopikrishnajha, @avbodas, and the OpenVINO community,

My name is Hyeongseob Jo. I recently completed my M.S. in Computer Engineering at Dongguk University, Seoul, South Korea, and I am currently working as a Junior Machine Learning Engineer (Intern) at ML2 (https://www.kc-ml2.com/).

I am very interested in participating in GSoC 2026, particularly in Project #34: Agentic AI for Predictive Maintenance of Cognitive Cities Critical Infrastructure.

Before preparing my proposal, I would like to briefly introduce my background and ask a few questions to better understand the expected scope of the project.

Background

1. During my graduate studies, I worked on a multi-illumination-based automated defect inspection algorithm for display panels in collaboration with LG Display. I later implemented my own prototype-based defect classification model (ProtoDC-Net) and evaluated it on public datasets.

• Project repository: https://github.com/johyeongseob/protodc-net

2. In September 2025, I participated in the Samsung Collegiate Programming Challenge (SCPC), where I developed a multimodal VQA model.

• Project repository: https://github.com/johyeongseob/scpc2025-jhs

3. Currently, at ML2, I am working on deploying a MobileVLM model on iOS devices using Xcode.

• Related presentation: https://www.notion.so/kc-mic/MobileVLM-321c20ac3b03806e958bcca417ee5db0?source=copy_link

4. I also enjoy reading and reviewing research papers. I maintain a personal blog where I summarize papers I have studied (most posts are in Korean).

• Blog: https://johyeongseob.tistory.com/category/%5BPaper%20Review%5D

5. My resume can be found here:

• Resume: https://github.com/johyeongseob/cv-johs/blob/master/cv_johs.pdf

OpenVINO Contribution

I have also submitted a pull request to openvino_contrib:

1. PR: docs: add offline dependency installation instructions openvino_contrib#1047

---

My Understanding of Project #34

From the project description, my understanding is that the goal is to leverage the predictive maintenance pipeline provided in Intel Metro AI Suite 26.0 to build a sample agentic AI application for monitoring critical infrastructure. The system would likely use LLMs and VLMs optimized with OpenVINO, and the final application would be validated on Intel Core Ultra hardware using open datasets.

Before drafting a formal proposal, I would appreciate clarification on a few points.

1. The scope of an agentic predictive maintenance pipeline seems quite broad.
   Are there specific domains or datasets that you would recommend for this project (e.g., bridges, utilities, solar panels), or is the contributor expected to select an appropriate open dataset?

2. From the project description, I assume that multimodal models (e.g., VLMs) may be used to analyze visual infrastructure data together with textual reasoning.
   Should the project focus more on vision-based inspection models, or on the agentic AI reasoning layer built on top of existing models?

3. The description mentions LLMs and VLMs optimized with OpenVINO.
   Should the project primarily reuse these optimized models, or is additional model optimization expected as part of the project deliverables?

Thank you for your time and guidance.
I look forward to learning more about this project and preparing a proposal. :)

Best regards,
Hyeongseob Jo

[johyeongseob]

Hi, following up on this discussion.

I recently built a small demo that might be relevant here:
https://github.com/johyeongseob/metro_mini_agent

The idea was to implement a lightweight agentic pipeline for predictive maintenance using a Vision-Language Mode (I used Qwen-2.5-VL-3B), without relying on heavy orchestration frameworks.

The pipeline is structured as:
Perception → Structured JSON → Risk Scoring → Decision → Action

For a single infrastructure image, it produces outputs like:

{
"perception": {
"defect_type": "Infrastructure Defect",
"severity": 0.5,
"confidence": 0.8,
"location": "bottom-right",
"recommended_action": "Immediate inspection and repair"
},
"risk_score": 0.4,
"decision": "inspect",
"action": "Schedule inspection"
}

[hmoustaf]

Thank you for the follow up johyeongseob.
Great to see your interest and active follow up.
Please see the answers I provided to the 3 questions you had (answers in my last comment). I provided a resource for an existing pipeline we released and would be good starting point. You can sure expand with the work you are developing now.

[hmoustaf]

Thank you for your interest johyeongseob.

I am answering your 3 questions in your first comment:
Question (1) The scope of an agentic predictive maintenance pipeline seems quite broad.
Are there specific domains or datasets that you would recommend for this project (e.g., bridges, utilities, solar panels), or is the contributor expected to select an appropriate open dataset?
Answer: Yes we will choose one or two use cases and we will guide on open data set to use.

Question (2): From the project description, I assume that multimodal models (e.g., VLMs) may be used to analyze visual infrastructure data together with textual reasoning.
Should the project focus more on vision-based inspection models, or on the agentic AI reasoning layer built on top of existing models?
Answer: Multimodal is expected to consider. We have released an agentic pipeline now available here (https://docs.openedgeplatform.intel.com/2026.0/edge-ai-suites/ai-suite-metro/app-blueprint-agentic-maintenance.html) that is expected to be good starting point.

Question (3): The description mentions LLMs and VLMs optimized with OpenVINO.
Should the project primarily reuse these optimized models, or is additional model optimization expected as part of the project deliverables?
Answer: we will have as first priority the models that are enabled with OpenVINO.

[hmoustaf]

Did you confirm your choice of this project?

[johyeongseob]

Hi @hmoustaf,

Nice to meet you, and thank you for your detailed answers! 🙏

Yes, I confirm that this is my primary project choice.

I am currently preparing my proposal based on the agentic pipeline you shared and my recent work.

As the deadline is approaching, I plan to submit my proposal within the next 24 hours.

Thank you again for your guidance.

[hmoustaf]

Great! We look forward to your propoal.

[johyeongseob]

Thank you! 😊