Notes
Short records that do not need to become articles. 2 in total.
A summary of the CSDN period — Python, NumPy/pandas, data analysis, visualization, and a few competition projects first; then transformers, LLMs, and agents (tool calling, the agent loop, multi-agent setups, prompt / context / harness, Claude Code, prompt-injection defences, agent checkpoints). From September 2026 the blog lives on my own site; the old CSDN posts stay up but are no longer updated.
A debrief of two or three interviews at small companies. The projects were one agent and one LLM, and the questions landed in the same two places — why GQA works (MHA's dimensional expressiveness, MQA's KV cache, GQA as the middle ground), three layers of prompt-injection defence (model / agent / permissions), VRAM and CPU-and-memory contention with a hundred-plus concurrent users, what to do when a sub-agent gets stuck (timeouts and a scheduler), which fields a trace should observe and the ten questions it must answer, recovering information lost in summary compression, what a high-risk approval prompt has to show, and saving checkpoints by event rather than per step.