Notes, essays, and research writing.
A first-principles walkthrough of building a memory subsystem for an LLM-driven agent. It covers why a single store cannot satisfy latency, durability, and recall simultaneously, and how a dual-layer design resolves the trade-off.
A walkthrough of building a dual-branch graph attention network for sensor-based fault diagnosis: constructing spatial and spectral graphs, running parallel attention, and fusing the branches without letting one dominate the other.
A first-principles walk through hyperspectral image classification. It covers why the data shape breaks ordinary CNNs, how spectral attention recovers the inductive bias, and what an external memory module of class prototypes actually contributes.
A walkthrough of a transfer-learning pipeline for imbalanced medical image classification: backbone selection, fine-tuning regime, augmentation policy, loss engineering, and the evaluation protocol that actually measures what you care about.