06 / ENGINEERING NOTES
Engineering Notes
Long-form notes on AI, engineering systems, and organizational evolution.
Why I Put OCR Before the LLM: Image Understanding Is More Than Uploading a File
A production account of how Shixiseng OCR Service validates, preprocesses, rasterizes, recognizes, and structures images and scanned PDFs before sending reliable context to an LLM.
The Next AI R&D Organization: From Functional Roles to Project Cells
AI will not erase expertise, but it will weaken fixed role boundaries. The basic delivery unit can shift from functional queues to small project cells with the capability to validate and own market outcomes.
Software Is Becoming Disposable: How AI Rewrites SaaS and Internal Systems
As software becomes cheaper to generate and rewrite, companies no longer need a permanent SaaS product for every local need. More internal software will be contextual, short-lived, and continuously rebuilt—and startups will need revenue earlier.
One Content Source, Three Reading Experiences: Web, WeChat Mini Program, and macOS
How I turned one source of technical notes into a responsive website, a WeChat Mini Program, and a native macOS reader through separate read-only APIs and platform-specific experiences.
A Vertical AI Model Is a System: Building Atlas, a Local AI Career Assistant
How I combined Qwen3-0.6B, MLX LoRA, a local job index, resume parsing, deterministic rules, intent routing, and release gates to build Shixiseng Atlas.
Building CoreX: A Native Real-Time macOS Monitor in Swift
How CoreX combines Mach, IOKit, Metal, SwiftUI, AppKit, and MenuBarExtra to monitor CPU, GPU, memory, disk, network, and thermal pressure entirely on the Mac.
FaceFizz: Building a Privacy-First Camera Inside the Browser
An engineering retrospective on FaceFizz, from camera permissions and real-time Canvas effects to local MediaPipe segmentation and explicit resource cleanup.
A Packaging Bug I Found in Homebrew Redis 8.10.0
A short debugging note on a Redis startup failure caused by enabled module directives whose libraries were missing from the Homebrew bottle.
Building a Recruitment LLM from Zero to One: A BOSS Zhipin–Style Case Study
A practical open-source blueprint for a recruitment-domain language model: product goals, data governance, continued pretraining, SFT and DPO, RAG, candidate–job matching, evaluation, safety, deployment, and a 12-week delivery plan.
The Moat in Large Language Models Is Not Just Code—It Is Data
As Transformer implementations, training frameworks, and open weights become widely available, the scarce advantage shifts toward data quality, data mixtures, verification, governance, and the feedback loop that improves them.