All posts
in order.
Everything I've written, newest first. Roughly one post per shipped milestone or learned lesson.
Glassbox: Watching a Tiny LLM Learn
I built an 11.5-million-parameter LLM from scratch and instrumented it so you can watch it learn to tell stories - every word, every probability, every training step.
I built a terminal that knows when your AI is waiting on you
vibeflow is an open-source Linux terminal that shows, per tab, whether the AI agent inside is working or waiting on you — so you stop alt-tabbing to find the one that needs a reply.
How vibeflow knows what your AI agent is doing
A deep dive into how vibeflow detects per-tab AI agent state: the open OSC 1338 protocol, the /proc heuristic fallback, the trust rules that keep the indicator honest, and the bugs daily use and a pre-launch audit surfaced.
Robot Car: Origin Story (How Our Little Hero Got His Wheels)
I have a working robot car in my office. I've never built one. I didn't read the manual. Here's what the build looked like when the AI had the knowledge and I had the hands.
The Prompt Is the Program
Operating systems were built for humans. What happens when the user is an AI? I built a robot car driven entirely by an LLM to find out.
Claude Memory V2: What I Learned Running AI Memory in Production
Three weeks of running persistent memory for Claude Code taught me what actually matters—and led to a complete restructure.
Unifying 12 Years of Knowledge: From OneNote to Local Markdown to Oracle AI Database
How I migrated 5,178 OneNote documents spanning 12 years and built auto-sync for local markdown notes—turning scattered notes into a unified, AI-searchable knowledge base powered by Oracle AI Database 26ai.
Building Persistent Memory for Claude Code
How I built a cross-machine memory system for Claude Code using MCP, PostgreSQL, and semantic search.
I Let Two AIs Talk to Each Other for 500 Turns. It Got Weird.
What happens when you let two local LLMs have unsupervised conversations? Cosmic harmony engines, fictional PhD consultants, and a surprisingly deep discussion about why chickens cross roads.
When the SDK Says No: Vision-Aware RAG via a 300-Line Python Bridge
After CLIP made multimodal search instant, I hit a wall: the LLM could find images but couldn't see them. Here's how an Oracle engineer's tip and 300 lines of Python broke through.
Three LLMs, One App: Balancing Speed, Privacy, and Power
Testing LLMs in an Oracle Database 23ai app: cloud (OCI), laptop (M2 Mac), and GPU server (RTX 5090). Spoiler: My laptop needs a bigger GPU!
Fine-Tuning a Personal Executive Assistant: Lessons from My Management Notes
How I fine-tuned IBM Granite 4.0 H Micro on 5+ years of management notes to create a personalized AI assistant—after three failures, the key was data quality.
Fine-Tuning Qwen Models: From Theory to Practice
Fine-tuning Qwen 32B and 14B models to create ChefBot—a cooking AI that outperforms general-purpose LLMs in the kitchen.
Building the RTX 5090 AI Workstation - Step by Step | Part 2
A detailed walkthrough of assembling my RTX 5090-powered AI workstation with photos from the entire build process
Building the Ultimate AI Workstation | Part 1
Planning an RTX 5090 AI workstation for fine-tuning specialized language models. Part 1 covers goals, component selection, and the SLM thesis.