Transforming operational scheduling into a system-level intelligence process. We move beyond task-level feasibility to optimize for risk awareness, long-term efficiency, and structural resilience.
Why traditional scheduling pipelines miss the bigger picture.
Current pipelines prioritize short-term rule satisfaction. This leads to "invisible" long-term risks like overdue accumulation, regional congestion, and uneven workload distribution that only emerge over time.
A "World Model" that treats scheduling as an interconnected system. We evaluate candidate schedules not just on validity, but on probabilistic risk, duration deviation, and resilience.
Constructed a heterogeneous graph from the CLICK dataset: 5.4M task nodes and 22k engineer nodes, linked by operational logic to capture hidden dependencies.
Leveraging GNNs to capture structural and temporal dependencies.
Non-graph baseline treating tasks independently. Used to benchmark the value of structural information.
Gradient boosting tree model designed for tabular data. Serves as a strong traditional baseline to compare against neural network and graph-based approaches.
Relational Graph Convolutional Networks. Distinguished semantically different interactions (e.g., assignment vs. geographic).
Scalable Inductive Learning. Aggregates neighborhood features to capture local structural context efficiently.
Heterogeneous Graph Transformer. Uses attention mechanisms to weigh messages based on node and edge types.
Empirical evidence of system-level learnability.
Smooth L1 Training Loss across GraphSAGE, RGCN, and HGT models.
Comparison of Evaluation Loss across MLP, LightGBM and GraphSAGE.
t-SNE of hidden states reveals a structured representation space where similar task types, districts, and departments cluster, with completion time forming a vertical gradient.
From predictive modeling to practical insights and operational visibility
A web application that provides insights in task completion time and operational risk.
User uploads a set of work order schedules to begin graph construction and model inference.
Dashboard shows high-level overview of key scheduling metrics (task completion time, workload imbalance, district-level performance, etc).
Per-assignment completion time predictions displays in a sortable table. Users can export data in CSV format for further analysis.
Chart visualizations show performance distributions by engineer, department, and district, revealing patterns in operational risk.
Users can interactively explore the heterogeneous graph, visualizing relationships among nodes (e.g., assignments, tasks, engineers) alongside predicted risk values.
A demonstration of our scheduling intelligence platform.
Note: Data shown in the images and demo are artificially generated for demonstration purposes.