Writing

Build logs

Build logs from shipping MacGet and AI systems in public — the engineering, the trade-offs, and what broke along the way.

Latest

Signal Boy: a zero-dependency WiFi survey tool you run with npx

I wanted to find the dead zone in my flat without installing a heatmap app or signing up for anything. So I built Signal Boy — one npx command, a live dBm readout on a Game Boy LCD, and no dependencies at all. Here's what that constraint cost me.

signal-boynodejswifiAug 7, 2026 · 6 min read

MCP servers vs agent skills: the payload is the integration

Shipping an MCP server is the easy half. The work that decides whether an agent can actually drive your product is payload design — closed vocabularies, schema-constrained returns, and tool descriptions written for a reader who will not ask a follow-up question.

mcpmodel-context-protocolagent-skills

Aug 3, 20266 min read

MacGet 1.3.0: BitTorrent, book catalogs, and deleting my own adaptive concurrency

MacGet 1.3.0 adds BitTorrent and OPDS book catalogs to the macOS download manager — and rips out the adaptive connection ramp I shipped as a headline feature in 1.1.0, because the heuristic read a saturated link as a failed experiment.

MacGetSwiftmacOS

Jul 30, 20267 min read

Automating blog cover art with Gemini 3 Pro Image: my 'no legible text' rule was the bug

I wrote a Node script to generate whiteboard-sketch covers for every post on this blog, and banned the model from drawing readable text. Nano Banana Pro ignored the ban, spelled everything correctly, and proved the constraint was a leftover assumption from 2023.

gemininano-banana-proai-image-generation

Jul 26, 20265 min read

Claude Opus 5 vs Fable 5: how I'm routing nodes in a LangGraph agent

Opus 5 landed at half Fable 5's price and claims to approach it on most tasks. For a multi-agent research desk the real question isn't which model wins — it's which node gets which, and the two API changes that will silently break your existing graph.

claude-opus-5claude-fable-5langgraph

Jul 26, 20266 min read

Run a build-log blog on your portfolio for $0: Next.js, Supabase, Redis, and Resend on free tiers

How I added a full build-log blog to my portfolio — likes, email subscribers, and new-post notifications — without paying for anything. The stack, the code, and the one Vercel Hobby limit that broke my cron.

nextjsbuild-in-publicsupabase

Jul 22, 20266 min read

How we're architecting a financial AI analyst: a multi-agent research desk on LangGraph

The system behind an AI that reads filings and answers like an analyst — specialized agents on LangGraph over a Vertex multimodal RAG data plane. What the architecture looks like, and the four things that broke on the way there.

financial-aimulti-agentlanggraph

Jul 16, 20268 min read

Hard gates and soft gates: stopping system prompt leaks in LangGraph agents

Prompt injection isn't one problem, so it doesn't get one fix. How I layered deterministic hard gates and an LLM-as-judge soft gate across a LangGraph agent — and why the refusal itself turned out to be the leak.

prompt-injectionlanggraphlangchain

Jul 15, 20267 min read

Agentic RAG for financial research: grounding agents on Vertex AI's multimodal RAG Engine

LangGraph runs the control plane; Vertex AI's managed multimodal RAG Engine runs the data plane. How I got retrieval quality and groundedness where I needed it — because in financial filings, the number lives in the chart, not the paragraph.

agentic-raglanggraphvertex-ai

Jul 12, 20267 min read

How I actually code with Claude Code: /clear, /compact, and a CLAUDE.md that earns its keep

The context-management habits that keep an AI coding agent sharp over a long session — when to /clear, when to /compact, and how to write a CLAUDE.md the model actually reads.

claude-codeai-assisted-devworkflow

Jul 10, 20267 min read

System design notes for AI products: fast, cheap, and boring

The infrastructure lessons behind running LLM features in production — streaming, caching, cost routing, and treating the model as an unreliable network call.

system-designai-agentsbackend

Jul 10, 20265 min read

How I architect LangGraph agents that actually ship

The structure I reach for when building production AI agents — explicit state, small nodes, tool boundaries, and knowing when NOT to let the model decide.

ai-agentslanggraphsystem-design

Jul 8, 20266 min read

Building MacGet's download engine: chunks, actors, and yt-dlp

A build log on how MacGet downloads files fast on macOS — chunked HTTP parallelism, an actor-based engine in Swift, browser-capture over Native Messaging, and bundled yt-dlp for video.

MacGetSwiftmacOS

Jun 27, 20267 min read

Get new build logs in your inbox

New engineering write-ups on MacGet and AI systems. No spam, unsubscribe anytime.