06 / ENGINEERING NOTES
Engineering Notes
Long-form notes on AI, engineering systems, and organizational evolution.
Do Large Language Models Understand? From Next-Token Prediction to the Boundary of Intelligence
Large language models are built from next-token prediction and matrix operations. Does that mean they understand nothing? A careful distinction between representation, functional understanding, consciousness, and human experience.
Read articleAgent Architecture for SMBs: A 90-Day Path to Production
Do not copy an enterprise AI platform. Build a production agent with deterministic orchestration, narrow tools, least privilege, evaluation, and measurable business outcomes.
Building Docket with Pragmatic DDD
A code-based look at how Docket uses bounded contexts, domain services, and a modular monolith to manage a professional-services workflow—and how the architecture can evolve.
Xinxu Cloud Brain AI: Building a Reliable Multimodal Emotion System
A production architecture for voluntary text, speech, and drawing inputs, combining modality-specific encoders, robust fusion, multi-task learning, calibrated uncertainty, privacy governance, and continuous evaluation.
Modern Concurrency: From Execution Models to Production Reliability
A practical guide to memory models, Java virtual threads, Go channels, Python asyncio, structured concurrency, backpressure, cancellation, and observability.
LangChain vs. LlamaIndex: Choosing a Framework and Building Production RAG
A practical comparison of LangChain, LangGraph, and LlamaIndex, followed by an end-to-end production RAG architecture covering ingestion, hybrid retrieval, reranking, citations, evaluation, and access control.
How to Build an AI-Native Team
AI-native transformation is not a tool rollout. It is a redesign of goals, workflows, human-agent responsibilities, organizational context, evaluation, and governance.
From Query Understanding to Knowledge-Driven Recommendations
A production-minded architecture for tax and finance search, combining structured query understanding, hybrid retrieval, knowledge governance, reranking, and continuous evaluation.
My AI Programming Workbench: Orchestrating Models Instead of Hunting for One Best Tool
How I combine Ollama with Qwen3.5:9B, OpenCode, Claude Code, Codex, and desktop GPT through project context, architecture review, task decomposition, cross-review, and evidence-based quality gates.
Building Production Intelligent Risk Control: Streaming, Lakehouse, Rules, and Models
A production architecture for low-latency, replayable, auditable, and degradable risk decisions using Kafka, Spark Structured Streaming, CDC, Hadoop/Iceberg, rules, and models.