Articles
In-depth articles on AI, web development, cloud native, and indie hacking.
"Model leaderboards cannot replace business evaluation. This article designs a five-layer LLM application evaluation framework spanning evaluation-set construction, rule-based metrics, model judges, human review, and production observation to support model selection, Prompt regression, and RAG optimization."
Agent tasks for research, procurement, and batch data processing often span minutes or even days. This article explains how to persist task snapshots, restore checkpoints, handle timeouts, and avoid duplicating external side effects after restarts.
A single Agent can confidently go off course on complex tasks. This article explains three multi-Agent orchestration patterns (supervisor, hierarchical, and handoff), implements a "supervisor + specialists" system with LangGraph's state graph, and covers the state, loop, and human-intervention engineering needed for production.
As memory grows, you can't save tokens by overwriting old summaries. This article designs fixed input snapshots, derived lineage, key-fact fidelity tests, invalidation propagation, and reversible switching.
Taking over an unfamiliar codebase used to take a week of reading; an AI agent can map the skeleton in hours. Learn to use Claude Code/Cursor to onboard fast: architecture overviews, call-chain tracing, feature explanations, risk identification, and good codebase indexes (AST/docs).
Deleting one row in the primary table won't clear vectors, summaries, caches, or restored backups. This article uses a deletion state machine, generational tombstones, derived lineage, and end-to-end canary verification to truly forget.
Not every LLM task needs to run in real time. Batch classification, summarization, labeling, and content generation can cost half as much with Batch APIs. This article covers OpenAI and Anthropic Batch APIs, self-hosted batching with vLLM, job pipelines, and real-time vs. batch tradeoffs.
User preferences, project decisions, and roles all change. This article separates facts using valid time and recorded time, explicitly handles supersession, late-arriving, and concurrent conflicts, and avoids last-writer-wins producing incorrect memory.
Server Components power the App Router, but their boundary with Client Components is often misunderstood. This guide explains how RSC works, when to use 'use client', server-to-client data flow, caching, Suspense streaming, and common mistakes such as marking everything as client code or using browser APIs on the server.
Preferences and facts a model extracts from a conversation cannot become long-term memory directly. This article designs candidate proposals, deterministic validation, deduplicated commits, revocation lineage, and a contamination-rate gate.
Writing an interface does not mean you understand TS. This article covers conditional and mapped types, infer, template literal types, the limits of type gymnastics, and practical patterns for API response inference, type-safe routing, and props autocomplete, turning TS from “annoying to write” into a bug catcher.
Passing offline evaluation does not mean an agent is reliable in production. This article designs online success rate, human takeover, rejection quality, tool recovery, and cost metrics to help teams use layered observation to judge whether a version deserves more traffic.