CI Gates, Test Result Caches, and the Code Graph That Might Tell Us What to Test
What I learned from splitting Rails CI by risk, caching exact test results, and exploring whether AI and code graphs can make test selection safer.
AI-Native Engineering
AI that supports engineering judgment, with verification, privacy, and approval built in.
The problem
Teams want the leverage of coding agents and LLM features without handing over control, leaking data, or shipping output nobody verified.
What I offer
- Agent architectures with explicit context, tools, memory, and human gates
- Local and hosted model workflows sized to real hardware and real budgets
- Evaluation and verification loops so AI output earns trust
- Bounded AI features inside existing products rather than platform rewrites
Start here
- LLM Strengths and Limitations: A Practical Framework · A practical framework for deciding when an LLM is the right tool for a task, when it isn't, and how to architect around its limitations.
- AI Application Architecture: LLM + Memory + Tools · The five components a production AI application actually needs beyond the model itself: memory, tools, retrieved knowledge, and structured context.
- How an LLM Coding Agent Actually Builds Software · How LLM coding agents actually build software: the model, the tool loop, context management, precise patching, and verification against real tests.
All writing in this area (50)
The One-Person Company Is a Coordination Problem
I am exploring how a one-person company can use agents, observability, and market intelligence without turning automation into an ungoverned second company.
When Productivity Turns Against Its Purpose
The AI boom promises abundance, but without a fairer distribution of power it may deepen the poverty, dependency, and cultural depletion it claims to overcome.
When the Report Button Becomes a Competitive Weapon
Unfair competition has learned to exploit platforms, moderation systems, and the gap between takedown and appeal, from invented users to AI-generated consensus.
Hermes vs OpenClaw, NanoClaw, ZeroClaw, IronClaw: Agent Architecture
A practical, architectural comparison of Hermes, OpenClaw, NanoClaw, ZeroClaw, and IronClaw, and why a layered agent system beats picking one winner.
Could AI Simulate Billions of Possible Human Lives?
An exploration of large language models as compressed maps of human experience, and what a branching world model built on them might reveal.
Codebase Memory for Coding Agents: Graft, Graphify, and codebase-memory-mcp
A practical comparison of Graft, Graphify, and codebase-memory-mcp: what each indexes, how each feeds coding agents, and when to use one instead of another.
Testing Small Local LLMs on Modest Hardware: A Portable Workflow
A repeatable way to download, serve, benchmark, and choose small GGUF models for Hermes and coding-agent work across different CPU and memory budgets.
Local Gemma with Hermes: Cheap Delegation for Rails Maintenance
A practical guide to using Hermes Agent with a local Gemma model for routine coding work, while reserving cloud models for tasks that genuinely need them.
Agent 基建的三层爆发：电脑、运行时、技能方法论
从 Prime Agent、Cloudflare Computer、agent-skills、mattpocock/skills 和 Superpowers 看 Agent 基础设施正在如何从聊天框走向持久工作台、长任务运行时和工程方法论。
Golden Rules for Cheaper, Safer LLM Agents
Notes on turning repeated AI workflows into scripts, skills, cron jobs, and guardrails so agents become cheaper, safer, and more useful over time.
Loop Engineering and Graph Engineering: Working With LLMs
A practical reflection on prompt, context, memory, loop, and graph engineering, and what each layer reveals about working with LLMs day to day.
Life Is Not Git
A personal reflection on Life OS, LLM memory, and digital twins: why being perfectly archived and searchable is not the same as being alive.
Tightening Hermes Secrets with Bitwarden
A practical walkthrough for moving Hermes API keys out of local env files and into Bitwarden Secrets Manager, with verification and cleanup steps.
Don't Leave Good AI Workflows in Chat
A practical note on promoting repeated AI workflows into slash commands, scripts, and cron jobs, weighing judgment, cost, and safety trade-offs.
Shared AI Artifacts Aren't Private From a Hard-to-Guess URL
A security note on the Claude shared-chat and Artifacts leak, and why AI workspaces need clear boundaries between private, unlisted, and public content.
Voice Control for Hermes Agent over Telegram
How I wired Hermes Agent to accept Telegram voice notes, transcribe them locally, and reply with local text-to-speech without adding another paid speech API.
The Hard Part of AI PR Review Is Not Reading the Diff
A practical look at Alibaba OpenCodeReview, why diff-only AI review is not enough, and how to use AI reviewers without trusting them too much.
Tightening Hermes Agent Security for Local Sensitive Work
How I tightened my Hermes agent setup after adding a local model for PII and credential-sensitive work: tool limits, manual approvals, and a quiet watchdog.
Securing Hermes Agent with OpenShell
Why I run Hermes Agent inside OpenShell when it reads untrusted content, and how filesystem, network, process, and credential policy change the security model.
Durable codebase context for Hermes: GitNexus, LSP, AGENTS.md
How I gave Hermes a persistent local code graph, semantic diagnostics, and stable project instructions without turning agent memory into a stale symbol index.
Giving Hermes Read-Only Access to GCP, AWS, and Kubernetes
A practical setup giving an AI agent read-only access to debug QA and production on GCP, AWS, GKE, and EKS without touching workloads or secrets.
Building an AI English Conversation Practice App
Notes on building an AI speaking practice app: STT, TTS, Gemini, ElevenLabs, latency, topic control, and useful feedback for oral English learners.
Cloud Run as a Gemini proxy for an unsupported region
How I used Google Cloud Run as a small AI gateway so clients in an unsupported region can still call Gemini through a supported GCP region.
Running Hermes on a Cloud Machine as a Daily Routine Assistant
How I wired Hermes Agent on a remote Linux box into Telegram, email, calendar, error monitoring, GitHub, and a daily morning briefing routine.
Tiny Puzzles for Testing and Debugging AI Agents
Small probe tasks for testing local AI agents: function calling, skill awareness, prompt-length pressure, cache behavior, and layer-by-layer debugging.
LM Studio Local Agent Runbook: Pi and OpenCode Step by Step
A step-by-step runbook for wiring LM Studio as a local OpenAI-compatible backend for Pi agents and OpenCode, with config examples and troubleshooting.
How an LLM Coding Agent Actually Builds Software
How LLM coding agents actually build software: the model, the tool loop, context management, precise patching, and verification against real tests.
Using LM Studio and Gemma as a Local Engine for Coding Agents
Using LM Studio and Gemma as a local OpenAI-compatible backend for coding agents: what broke, which settings mattered, and where it still falls short.
Building a Local Coding Agent (Codex/Claude Code Style) in Gemma
Building a local coding agent around Gemma 4 26B taught me the model wasn't the problem: my agent loop was, and fixing the loop is what made it work.
A Local, Multi-Tenant Graph RAG System with LightRAG
How a local Ollama and LightRAG stack turns Graph RAG into a cheap, multi-tenant retrieval layer that only sends precise context to cloud models.
Token Efficiency for AI Coding Agents in 2026
Six practical techniques for cutting token spend in AI coding agents, from skeleton indexes to symbol-level retrieval and semantic caching.
What Subagents Are Actually Good For in Codex
A field guide for AI-heavy early adopters on when subagents in Codex speed work up, when they slow it down, and how to use them without creating chaos.
Git Worktree for Parallel AI Development
How to use git worktree to run parallel AI coding sessions safely, with practical patterns for isolating gems, databases, ports, and runtime state.
MCP, Skills, Agents, Rulesets, and Hooks: A Practical Map
A practical introduction to the main building blocks of modern AI development, with diagrams that show how MCP, skills, agents, rulesets, and hooks work together.
Intelligent Applications and the Personal OS
The closing piece of a 12-part series on how software is shifting from apps you click through to intelligent systems and a personal OS that act for you.
AI Employees: Building Your Digital Workforce
What changes when AI moves from a tool you operate to an autonomous agent you manage like an employee, and how to do that responsibly.
I Let an AI Agent Write Code for a Week. Here's What Actually Worked.
A week of testing OpenHands, the autonomous AI coding agent, on real projects: what worked, what failed, and the setup that made it reliable.
The One-Person Company: AI as Force Multiplier
How AI lets a single founder cover work that used to require a small team, and the real trade-offs of choosing to run a company alone.
The AI-Era Developer: From Coder to AI System Architect
What a developer's job actually becomes when AI can write the code: which skills stop mattering and which ones become the real differentiator.
AI's Impact on the Software Development Lifecycle
How AI is changing requirements, design, coding, testing, deployment, and maintenance, and which parts of the software lifecycle stay stubbornly human.
From Logic-Driven to Context-Driven Software
How software built by encoding explicit rules is giving way to software built by shaping context, and what that changes about the engineering job.
Context Engineering: The New Software Engineering
Why reliable AI behavior comes from designing the whole context, not tuning a single clever prompt, and the five layers that make up good context.
Why AI Applications Are Ecosystems, Not Software
Why AI applications behave like ecosystems shaped by feedback loops and user adaptation, and what that means for how you design and monitor them.
AI Application Architecture: LLM + Memory + Tools
The five components a production AI application actually needs beyond the model itself: memory, tools, retrieved knowledge, and structured context.
LLM Strengths and Limitations: A Practical Framework
A practical framework for deciding when an LLM is the right tool for a task, when it isn't, and how to architect around its limitations.
Generalization: Why AI Looks Smart
How LLMs generalize from patterns instead of understanding, where that generalization holds up in practice, and where it quietly breaks down.
What is LLM: The Nature of Language Models
What large language models actually are: probabilistic token predictors, not databases or reasoning engines, and why the distinction matters.
Self-Hosted Whisper: A Speech-to-Text API with Ruby and Kubernetes
How to run OpenAI's Whisper as a self-hosted speech-to-text REST API with Docker, a Ruby client, and a Kubernetes deployment for transcribing audio.
Whisper: An Open-Source Alternative to Cloud Speech-to-Text
OpenAI's Whisper is a self-hostable alternative to Google's Speech-to-Text API, with a Flask wrapper and a C++ port covered here alongside the original model.