Writings
Back to homepage
Everything I've published.
Essays and event recaps on building with agentic AI.
Who Holds the Loop
I spent a week running GPT models inside Claude Code. Here is the harness that decides everything, OpenAI's new plugin, and the receipts from every claim I tested.
What Documentation Does an Agent Need? 729 CLAUDE.md Trials
729 blind trials on CLAUDE.md and skill files. The same document scored 87% and 13% depending on nothing but the filesystem underneath it.
What Form Should Knowledge Take?
Frontier AI teams are converging on what curated knowledge should look like as a system. RAG and retrieval get a new, smaller job: serving it.
How I Think About Knowledge in Finance
Finance standardized its data for two decades but rarely kept the reasoning behind decisions. AI makes that reasoning cheaper to find. How firms can manage it.
Building OptiMind: Memory Architecture for Claude Code on Mobile
Agentic memory is a harder problem than it looks. Here's what I learned building a memory-enabled agent around GitHub and Claude Code.
Let the Flywheel Spin on Top of the Translation Layer
What 6+ years of building and shipping, and 40+ AI rooms, taught me about earning trust in AI.
A Science-Based Skincare Routine
People keep asking what skincare products I use. My daily routine, paired with the dermatology and photobiology research that justifies each step.
Tracing the Minds Behind Claude Code
I spent a weekend reading two Claude Code source trees side by side, 13 months apart, to find what changed and what stayed the same.
Agents Don't Need Better Models. They Need Better Infrastructure.
I watched 300 builders hit inference limits at Mistral's worldwide hackathon. Three weeks later, GTC 2026 announced the infrastructure to fix it. Here's what agent developers need to understand about the serving layer beneath the models.
What Senior Business Leaders Actually Ask About AI
Ten questions Directors and C-suite leaders asked when I guest-lectured at Columbia's executive AI program: liability, data, switching costs, talent.
How I Learn Anything with AI
The feedback loop that used to require study groups, office hours, and weeks of iteration now happens in real time. How arguing with Claude turned six weeks of Causal Inference into a repeatable system for learning anything technical.
Connecting Local to Global: Inside Mistral AI's Worldwide Hackathon
7,000+ applications, 1,000+ builders, 7 cities, $200K+ in prizes: inside Mistral AI's first worldwide hackathon, from the judge's seat at the New York edition.
The AI Superpower for PMs: Context Engineering
Context engineering (not prompt engineering) is the PM superpower that separates AI-native product managers from everyone else.