Applied AI engineering — deep dives on software, machine learning, and the systems behind them.
In-depth explorations of software engineering, deep learning, and the technologies that power modern systems.
Featured
深入拆解：如何为 LLM 智能体组装上下文
Every agentic CLI silently decides what to send the LLM each turn — the whole transcript, just the recent K, a summary of the older parts. Same task, different decision, wildly different cost. We benchmark four representative context-assembly strategies on a real bug-fix fixture, see where catastrophic loops come from when caches die, and walk through how Claude Code, Codex, and opencode each navigate the same trade-offs.
构建 Model Armor：面向 LLM 的多层安全过滤
Google's Model Armor wraps gen-AI calls in a managed safety pipeline you flip on with a flag. This article rebuilds that pipeline from scratch — rule-based filters, classifiers, LLM judges, prompt rewriting, and output moderation — so you understand what each layer actually buys you and where the seams are when you swap pieces into your own stack.
如何加固一个你并不完全了解的 Web 应用
You inherited an app and can't vouch for what's inside it. Rewriting takes quarters you don't have. This is a layered playbook for hardening it from the outside — secret scanning, a WAF at the edge, bot protection, OSV-Scanner for dependencies, then SAST and DAST — in the order you should add them, and what each layer actually catches.
读懂 CNN——卷积、特征图与池化
What a convolution actually computes, and why stacking them lets a network see edges, then textures, then objects. Built bottom-up with interactive demos: feature maps you can poke at, pooling shown as the dimensionality squeeze it really is. By the end every diagram in a CNN paper should read like English.
神经网络是怎么学的：深入反向传播与梯度下降
Most tutorials wave at gradient descent and skip backprop, or bury it in chain-rule notation. This one builds both from the bottom up: what a gradient is, why the loss surface curves, and how the chain rule walks errors backward through the network — with widgets you can scrub to see each piece move.