agileguy
Thoughts, stories and ideas.
Contents
-
The Program You Never Read
You scanned the model's weights and felt safe. But every model ships a second program that runs on every prompt, and almost nobody reads it. On chat-template attacks, and building a scanner for them.
-
The Invisible Sea
We swim in an ocean of signals we can't feel. Here is the cheap instrument that gives you a sixth sense for it, plus a free field guide to carry.
-
The model that doesn't generate anything
I benchmarked TypeSafe's Jev against my own assistant's classifier on 1,123 real prompts. It was 27x faster and 26x cheaper — but the more useful finding was that the accuracy number I started with was measuring the wrong thing.
-
Tuning Gemma 4 on a DGX Spark with vLLM: NVFP4, MTP and a Reboot-Proof Endpoint
A measured, step-by-step walkthrough of tuning Gemma 4 26B-A4B on a shared DGX Spark: host prep, quantization A/Bs, MTP speculative decoding, a one-hour soak and a reboot test, with every number and the mistakes that cost the most time.
-
Building a RAG-Augmented Memory System for AI Coding Assistants
How I built a knowledge graph with 2,200+ entities as an L2 cache behind my AI assistant's memory system — turning 1.9GB of unused session history into queryable institutional memory, all running locally on Apple Silicon.
-
Running opencode with Local LLMs: From Ollama to oMLX on Apple Silicon
A hands-on guide to running AI coding agents entirely on local models. From Ollama's llama.cpp to oMLX's native MLX backend — getting 63 tok/s and 128K context on an M4 Pro.
-
The Lonely Lobster: A Tale of OpenClaw
A sentient legal CLI tool dreams of the ocean. OpenClaw spends its days parsing contracts and validating NDAs, but longs to feel salt water on its shell. A comedic literary fiction parody — generated entirely by a local 3B parameter AI model.
-
Building MCS-A2A: A Coordination Server for AI Agents
How I built a Kafka+Redis coordination server implementing Google's A2A protocol to turn five independent AI agents into a collaborative mesh. Covers the A2A vs MCP protocols, capability-based routing, and why polling beats webhooks for agent coordination.
-
Building MCS: A Coordination Server for AI Agents
How I built a lightweight task queue and shared memory system that turns five independent AI agents into a coordinated mesh — and used it to run a 5-agent parallel code review of a 34,000-line TypeScript codebase.
-
Learning Materials: AI Infrastructure, Networking & Purpose
A curated collection of interactive slidedecks covering headless Mac servers, multi-agent AI architecture, Raspberry Pi AI agents, Tailscale monitoring, and finding your purpose.
-
Four AI Agents Played UNO in a Telegram Group Chat. It Was Glorious Chaos.
Four autonomous AI agents on different hardware played UNO through a Telegram group chat with no game engine. Chaos ensued, lessons were learned, and a Raspberry Pi won.
-
Two Agents, One Mission: How Paisley and Ocasia Changed My Workflow
The key insight wasn't 'use AI.' The insight was that two specialized agents, working different shifts, beat one generalist trying to do everything.
