Training Apple’s On-Device LLM: LoRA, QLoRA, and What Actually Worked
Apple’s adapter toolkit assumes a beefy GPU. I got LoRA training running on a free T4 and a Mac.
Home
Recent Posts
When Your AI Tools Can’t Talk to Each Other
Claude Desktop and Claude Code are separate worlds. Here’s an MCP server that bridges them.
The 3B Wall: What Apple’s On-Device LLM Can and Can’t Do in Your Shell
I wired Apple’s on-device 3B LLM into three zsh hooks and benchmarked multiple approaches to make it write correct shell commands.
Let the AI Out: Agents as a Control Layer
One prompt, zero code, and a team of agents built a BLE monitoring system that exercised real judgment.
Let the AI Out: Edge AI on a Microcontroller — From Zero to Inference in 90 Minutes
What happens when you point an AI agent at a microcontroller and say ‘deploy this ML model on the device’?
Turning Claude Code Sessions into HTML Replays
A CLI tool that converts Claude Code session logs into interactive, self-contained HTML replays you can share, embed, or host anywhere.
Let the AI Out: Same Hardware Tools, Every AI Agent
The same MCP servers — BLE, serial, debug probe — working across Claude Code, VS Code + Copilot, and Cursor. That’s the point of a protocol.
Let the AI Out: Giving the AI Agent a Hardware Debug Probe
What happens when the AI agent can halt the CPU, set breakpoints, flash firmware, and inspect memory directly?
The Governor Module Problem: Trust, Control, and AI-Generated Code
We learned to trust compilers. Can we learn to trust AI? The answer isn’t about the code — it’s about the controls.
Let the AI Out: BLE + Serial — Using Both Interfaces at Once
What happens when the agent can see both the wireless interface and the debug console at the same time?
Let the AI Out: Giving the AI Agent a Seat at the Serial Console
Serial is where embedded engineers live. What happens when the agent can see your console too?
Let the AI Out: Closing the Loop with Real Hardware over BLE
An MCP server that lets AI agents interact with real BLE hardware — closing the loop between reasoning and physical devices.
What Is MCP? A Practical History of How AI Learned to Use Tools
Models became useful not by getting smarter, but by gaining ways to act on the world. MCP makes that interface sane.