CWB_Project: A Collaborative Human-AI Project Planning Assistant

CWB_Project is an Agentic AI system designed to address the pain point in project management where decisions from meetings/emails cannot be synced to official planning systems. It converts unstructured communication into structured plan updates and uses a human-in-the-loop approval workflow to ensure traceability and human oversight of changes. This project participates in the "SJ Project Planner Agent" challenge of the Microsoft Code Without Barriers Hackathon 2026, with its core innovation lying in balancing automation efficiency and human control.

A long-standing problem in project management: decisions from meetings and emails cannot be synced to official tracking systems, leading to plans deviating from reality and team cognitive biases. CWB_Project was created to address this issue and participates in the "SJ Project Planner Agent" challenge of the Microsoft Code Without Barriers Hackathon 2026, aiming to bridge the gap between unstructured communication and structured plans.

Core Workflow: Inbox reception → Intelligent extraction of task elements (title, owner, deadline, status, dependencies) → Change classification (NEW/UPDATE/CONFLICT) → Draft generation → Human approval → Atomic submission. Key principle: No changes can enter the official system without human approval. Technical features include: Intelligent extraction of task-related information with source references; Three-layer change classification mechanism; Human-in-the-loop approval (individual/batch operations, conflict resolution, confidence-driven UX, evidence traceability).

Multi-dimensional views: Tracker panel (filterable table + color cues for overdue/upcoming tasks), Plotly-powered Gantt chart, emergency panel, change log. Bidirectional traceability: Task history changes, timestamp difference comparison, source evidence references. Baseline comparison and replay mode: Real-time plan vs. baseline comparison (highlighting owner changes, date shifts, status changes), time-based replay of meeting records' contributions to the plan.

Four-layer unidirectional dependency architecture: UI layer (Streamlit Web App) → PlannerService layer (workflow control) → PlannerAgent layer (LLM tools) + Repository layer (data access) → Azure PostgreSQL. Tech stack selection: LLM uses OpenCode Go (DeepSeek V4 Pro/Flash); Storage uses Azure Database for PostgreSQL; UI uses Streamlit + Plotly; Deployment uses Azure Container Apps; Language uses Python 3.12, etc.

Uses the official CWB_SJ dataset (CC0 license): tasks_master.csv (about 50 task baselines), meeting_notes.jsonl (first 10 meeting records), emails.csv (first 5 emails). Users can populate the demo via the "Load sample dataset" button. Online demo link: https://sjplanner-sm2026.calmbush-865eff5f.eastasia.azurecontainerapps.io/

Problems solved: Breaking information silos, responsibility traceability, efficiency improvement, risk management. Applicable scenarios: Agile development Sprint planning and review, cross-departmental progress synchronization, task extraction from customer communications, asynchronous collaboration for remote teams.

CWB_Project demonstrates a practical application of Agentic AI: enhancing human decision-making rather than replacing it. The architecture design (four-layer unidirectional dependency, Repository pattern) provides a template for similar applications, and its emphasis on traceability and auditing makes it suitable for compliant enterprise environments.