https://domlee.dev Build journals on AI delivery, agent infrastructure, and engineering for Japanese-enterprise contexts. en Fri, 26 Jun 2026 12:44:35 GMT https://domlee.dev/posts/advising-an-ai-agent-rollout https://domlee.dev/posts/advising-an-ai-agent-rollout Mon, 15 Jun 2026 00:00:00 GMT An architect's memo: the risks that actually bite when you put an AI agent over real enterprise data, a phased runbook to de-risk it, and how to translate the same architecture for the exec and the platform team. ai agents data architecture advisory https://domlee.dev/posts/agent-governance-japanese-enterprise https://domlee.dev/posts/agent-governance-japanese-enterprise Mon, 15 Jun 2026 00:00:00 GMT Deploying AI agents over enterprise data for a Japanese company has constraints a US-centric playbook misses: data residency under APPI, an approval culture that wants confirmation gates, and Japanese-language grounding. Notes from the context. ai agents data japan governance https://domlee.dev/posts/governance-as-an-agentforce-action https://domlee.dev/posts/governance-as-an-agentforce-action Mon, 15 Jun 2026 00:00:00 GMT sf-dev-ai runs governance checks over MCP. Here's the same check as a native Agentforce action — an allow-listed, user-mode Apex InvocableMethod — plus when to reach for Flow or LWC instead, and why the security has to live in code. salesforce agentforce apex security build-journal https://domlee.dev/posts/grounding-salesforce-agent-data360 https://domlee.dev/posts/grounding-salesforce-agent-data360 Mon, 15 Jun 2026 00:00:00 GMT The useful questions an architect asks an org can't be answered by a model alone — they need the org's real config and unified data as context. Here's a retrieval layer over Salesforce metadata + Data 360, built twice (LlamaIndex and LangChain) and compared honestly. ai salesforce rag llamaindex langchain data-cloud build-journal https://domlee.dev/posts/text-to-soql-that-wont-leak https://domlee.dev/posts/text-to-soql-that-wont-leak Mon, 15 Jun 2026 00:00:00 GMT Generating SOQL from natural language is easy. Doing it without leaking restricted data is the architect's problem. A two-layer field-level-security guard: constrain the prompt, then enforce on the output. ai salesforce nlp security build-journal https://domlee.dev/posts/sf-dev-ai-governance-mcp https://domlee.dev/posts/sf-dev-ai-governance-mcp Wed, 03 Jun 2026 00:00:00 GMT Salesforce shipped Agentforce Vibes and an MCP server for build/deploy/migrate. The governance lane was open. Here's what I built — tier-annotated tools, a proxy that composes with @salesforce/mcp, and five things that surprised me. ai mcp salesforce open-source build-journal https://domlee.dev/posts/architecture-from-workload https://domlee.dev/posts/architecture-from-workload Mon, 04 May 2026 00:00:00 GMT An agent skill that turns infrastructure design into a constraint-solving problem. Seven dimensions, ten formulas, a swappable catalog. Same workload in, same architecture out — citable to a formula instead of a hunch. ai infrastructure claude-code build-journal https://domlee.dev/posts/langgraph-reflection-paradox https://domlee.dev/posts/langgraph-reflection-paradox Sun, 26 Apr 2026 00:00:00 GMT Adding a self-reflection node to a LangGraph ticket-triage agent dropped escalations from 60% to 0%. It also took a sev1 cross-tenant data-leak from 'escalate' to 'finalize'. The math says reflection helps; the tail says be careful where you put it. ai langgraph vertex-ai build-journal https://domlee.dev/posts/multi-sdk-llm-notebooks https://domlee.dev/posts/multi-sdk-llm-notebooks Sat, 25 Apr 2026 00:00:00 GMT A side-by-side benchmark harness for the OpenAI and Anthropic SDKs. Same task, same prompt, same input — and a few cost surprises. ai developer-tools open-source https://domlee.dev/posts/agentic-delivery https://domlee.dev/posts/agentic-delivery Mon, 20 Apr 2026 00:00:00 GMT An internal AI product shipped in a month — a small product team, a director-mode engineer, and three other teams that quietly made it possible. ai claude-code build-journal developer-tools https://domlee.dev/posts/agent-messaging-moment https://domlee.dev/posts/agent-messaging-moment Mon, 23 Mar 2026 00:00:00 GMT OpenClaw has 200k stars. Anthropic just shipped Claude Code Channels. I built the same thing a year ago with bash scripts. Here's what all of it is pointing at. ai mcp build-journal developer-tools https://domlee.dev/posts/claude-code-channels https://domlee.dev/posts/claude-code-channels Mon, 23 Mar 2026 00:00:00 GMT Channels let you push events into a Claude Code session and relay permission prompts to a remote device. Here's how the API works and what it actually enables. ai mcp developer-tools claude-code https://domlee.dev/posts/building-agenthive https://domlee.dev/posts/building-agenthive Sat, 21 Mar 2026 00:00:00 GMT I wanted a control plane for AI agent teams — persistent rooms, structured decisions, visible autonomy. Here's how I built it. ai developer-tools build-journal agenthive https://domlee.dev/posts/ai-browser https://domlee.dev/posts/ai-browser Tue, 17 Mar 2026 00:00:00 GMT I wanted to test MCP discovery in practice, so I built a browser extension. It grew. Here's what I found. ai mcp browser build-journal https://domlee.dev/posts/mcp-discovery https://domlee.dev/posts/mcp-discovery Tue, 17 Mar 2026 00:00:00 GMT MCP servers are proliferating but there's no standard way to find them. A case for decentralized discovery — no portals, no registration, just a file at a known path. ai mcp protocol open-source https://domlee.dev/posts/sonde https://domlee.dev/posts/sonde Thu, 26 Feb 2026 00:00:00 GMT A cloud operations platform that builds a knowledge graph of your AWS environment and uses AI to reason about security, changes, and compliance. ai cloud-security aws developer-tools open-source https://domlee.dev/posts/building-envibe-3 https://domlee.dev/posts/building-envibe-3 Sat, 31 Jan 2026 00:00:00 GMT Shipping v0.2, finding a security bug in my own security tool, and what I learned building with AI. ai developer-tools open-source build-journal https://domlee.dev/posts/building-envibe-2 https://domlee.dev/posts/building-envibe-2 Sat, 31 Jan 2026 00:00:00 GMT From design to first commit. Building an MCP-powered permission layer for AI coding assistants with Claude Code. ai developer-tools open-source build-journal https://domlee.dev/posts/building-envibe https://domlee.dev/posts/building-envibe Sat, 31 Jan 2026 00:00:00 GMT How a concern about AI assistants reading .env files turned into a permission layer. The research, the design, and the first architecture. ai developer-tools open-source build-journal https://domlee.dev/posts/envibe https://domlee.dev/posts/envibe Sun, 12 Jan 2025 00:00:00 GMT Introducing envibe - a permission layer that gives AI coding assistants granular access to environment variables without exposing sensitive secrets. ai security developer-tools open-source