Notes, essays, and research writing.
Design notes from ongoing work on LLM-driven drone navigation: which decisions a language model should own, which it must never touch, and how to structure the boundary between high-level mission planning and onboard perception and control.
Why linear agent pipelines break down on real tasks, what changes when agents, tools, and memory become nodes in an executable graph, and the design decisions behind GraphBit's non-linear orchestration model.
A retrospective on a line of work in hyperspectral image classification: why the data shape defeats ordinary CNNs, what spectral-spatial residual attention recovers, where dynamic kernels help, and why energy-based generative models with morphological attention were the natural next step.
Design notes on treating a photovoltaic inverter as a graph rather than a bag of sensor channels: how to build the graph, why a single attention view is not enough, and what a dual spatial-spectral graph attention network actually learns.
Why convolutional networks struggle on raw three-phase fault waveforms, how time-series imaging recovers a representation they can exploit, and where self-attention earns its place in a transmission line fault classifier.