Zhenggang Wang

Dual-branch frequency-domain pansharpening network. The Global Branch applies a single learnable anisotropic Butterworth filter over the full feature map; the Local Branch uses Butterworth Token Interaction (BTI) for patch-level adaptive texture enhancement. All operations are element-wise on the FFT spectrum for O(N log N) complexity.

Formulated pansharpening as low-dimensional orthogonal sketch learning via optimization unfolding, significantly reducing computational overhead. Introduced the Legendre Spectral-Gating Sketch, a learnable orthogonal projection operator that dynamically constructs sketch subspaces from Gram covariance using Legendre polynomial recurrence and Schulz orthogonalization. Designed an Adjoint-Guided Data Consistency Compensation module together with explicit Null-Space Residual Recovery, yielding an interpretable degradation-aware reconstruction pipeline grounded in physical observation models. Achieved SOTA performance on multiple pansharpening benchmarks with substantially lower computational cost than existing unfolding-based methods.

Full-stack AI tutoring platform built around a ReAct Agent that dynamically dispatches to 10+ domain-specific Skills and 16+ MCP tools including RAG retrieval, web search, and code execution. Hybrid RAG pipeline (dense + BM25 + cross-encoder re-ranking), LightGBM cold-start user profiling, and three-tier memory system. Multi-user isolation via Docker Compose.