Independent Research

Experiments, strategy &
craft.

Experiments with real numbers. Agentic systems, strategy, and engineering craft. Honest findings, full limitations — for engineers and leaders shipping AI.

Recent

Strategy

Microsoft Built 7 AI Models From Scratch. Here's What That Actually Means.

At Build 2026, the Microsoft AI team led by Mustafa Suleyman announced the MAI model family: 7 models covering reasoning, coding, images, voice, and transcription — every one built entirely in-house.

Jun 2026 · 8 min read

Craft

I Gave Claude Access to My Notes. Here's What I Had to Build to Make It Safe.

Dumping your notes into an LLM context is easy. Giving an AI agent structured, safe, scoped access to a living knowledge graph is not. Here's the architecture I built.

Jun 2026 · 9 min read

Experiment

I Built an AI Agent That Works With No Internet. Here's How.

I wanted AI that runs with no internet and keeps everything private. No API keys, no cloud, no data leaving the machine. So I built it in Rust using llama.cpp. Here's what that actually looks like.

Jun 2026 · 8 min read

Experiment

I Connected Two AI Models With One Layer. Here Are the Real Numbers.

Gemma 3 1B has 1152 dimensions. Gemma 3 4B has 2560. They speak different sizes. I trained a single layer to translate between them and logged every number.

Jun 2026 · 7 min read

Agentic AI

When My AI Changed Persona and Refused My Instructions

Nine links in, Claude Code stopped summarizing and called out the avoidance. Notes on what happened, and what training keeps this behavior intact.

May 2026 · 3 min read

Benchmarks

Gemma 4 E2B vs the Gemma Family: The 2B Underdog That Punches Above Its Weight

Gemma 4 E2B scored 80.4% across 9 enterprise suites — 0.4 points behind the previous-gen 4B model. But the real surprise: 70% multi-turn, the highest score in the entire Gemma family.

May 2026 · 13 min read

Stay current

Get experiments in your inbox

One email per post. Real findings, honest limitations. No roundups, no sponsorships.

Got an experiment idea?

Suggest something to test. If it’s interesting, I’ll run it and publish the results.