CS Capstone Project 2026
A multi‑agent platform for public opinion analysis with stance-aware retrieval, multimodal video understanding, and structured debate coordination.
📁 View source code on GitHub
🏛️ HKU · Supervisor: Prof. Kai Han
📌 Project progress & deliverables
📄 Project Setup & Proposal
✓ Completed – Environment, API contracts, full proposal & literature review.
🔍 Query Agent (stance-aware)
✓ Completed – Multi‑perspective sub‑query generation, credibility scoring, diversity‑aware reranking. Methodology documented.
🎬 Multimodal Agent (LangGraph)
✓ Framework ready – Video keyframe extraction, captioning & sentiment fusion pipeline.
⚠️ Image recognition still under refinement – standalone image understanding being actively improved.
⚠️ Image recognition still under refinement – standalone image understanding being actively improved.
⚙️ Agent Coordinator & Debate
In development – Moderator/Monitor design, multi‑round debate simulation, integration ongoing.
🖥️ Frontend MVP
In progress – Real‑time log streaming, interactive dashboard; basic API integration.
📑 Report Agent & Final integration
Planned – Structured report generation, knowledge graph visualisation, multi‑format export.
💡 Current milestone: Query Agent ✅ · Multimodal Agent (video understanding works; image recognition being polished).
Next: complete AgentCoordinator debate loop and frontend deployment.
Next: complete AgentCoordinator debate loop and frontend deployment.
🔧 Known limitation: The multimodal agent currently struggles with accurate image recognition (e.g., standalone photos, visual sentiment). We are actively debugging the vision‑language pipeline to improve robustness. Video processing is fully functional.
👥 Team Member
Lin Wenkai
Team Leader
Student ID: 3036657005
Kong Zeyu
Student ID: 3036655461
Li Yewen
Student ID: 3036656752
Miao Mengyu
Student ID: 3036661082
Huang Yawen
Student ID: 3036658633
🎓 Supervising Professor · Dr. Kai Han | Department of Computer Science, The University of Hong Kong
🏛️ The University of Hong Kong (HKU) – Faculty of Engineering, Computer Science Capstone Project 2025–2026
⚙️ System methodology (brief)
Our platform integrates four specialised agents orchestrated by a debate‑driven coordinator:
- 🔎 Query Agent – Decomposes user intent into 4 stance dimensions (official, support, oppose, neutral) + credibility scoring + near‑duplicate filtering.
- 🎥 Multimodal Agent – Built on LangGraph; downloads short‑form video, extracts keyframes (OpenCV), generates captions via VLM, and fuses text+visual sentiment. ⚠️ Image recognition module under active refinement.
- 🗣️ Agent Coordinator – Monitor + Moderator + Reader: implements iterative debate rounds to cross‑validate findings and mitigate bias.
- 📊 Report Agent – Template‑based report builder with modular chapters, knowledge graph, and interactive export (PDF/HTML).
📌 Current implementation focus – The Multimodal Agent (LangGraph pipeline) successfully processes video keyframes and performs sentiment classification on captions.
The team is debugging the image recognition pipeline to improve accuracy for standalone images and visual sentiment cues.
🌐 Deployment & access
This page is hosted via GitHub Pages from the main branch of the project repository.
🔗 Repository: https://github.com/lyw912/CapstoneProject
🌐 Live webpage: https://roselia-penguin.github.io/COMP7705-capstone/
🌐 Live webpage: https://roselia-penguin.github.io/COMP7705-capstone/
✅ All progress updates (methodology write‑ups, agent code, preliminary results) are regularly committed to the GitHub repository. Mentors can monitor development through both the repository and this dashboard.