Writing

On RAG, agents, AI in production, and things that break.

Latest2026-09-05

Looki Doesn't Need to Be an Agent. It Should Be an Agent's Eyes.

Auto-generated diaries are rarely worth revisiting, while proactive reminders often arrive too late. Looki's real role may be to turn life off-screen into observations you can inspect, correct, and give to any agent.

6 min read · AI · Agent

2026

The 128GB AI Beast Has Arrived — and Local AI Is Finally Moving In

The truly interesting part of NVIDIA and Microsoft's AI PC isn't that local models got smarter — it's that the agent is finally being managed as a process with permissions.

06-01 · 5 min read · AI · Agent

Just Edit the Paragraph

I figured adding an edit feature to my agent would be trivial. Instead I fell into three rabbit holes — str_replace, fuzzy patch, and Fast Apply. No best practice, just A/B tests.

05-22 · 4 min read · Agent · LLM · Engineering

SaaS Midlife Crisis: In-House Agents Flop, and We're Too Proud to Use Someone Else's

An observation from SaaStr — everyone's racing to build agents in the exact same way, keeping users locked in to feed context. Most SaaS can't pull it off, but they won't become the tool either.

05-21 · 3 min read · AI · Agent · Engineering

The MBTI Guide to Making Slides with AI

Five paths to AI-generated slides, each one a near-perfect fit for a different personality type. The worst AI can do is make the slide. The best it can do is write the code that renders it.

04-25 · 4 min read · AI · Productivity

What It's Like to Feed Yourself to an AI

I've been logging every hour of my day for seven years. Then I fed all of it to an AI.

03-31 · 3 min read · AI · Productivity

Most of Agent Engineering Has Nothing to Do with AI

The hardest part of building agents isn't prompt engineering — it's context management. When dynamic tool loading and skills both hit dead ends, sometimes you just build the ugly thing that works.

03-26 · 2 min read · Agent · AI · Engineering

I Was Accidentally Donating to Cloud Providers

I built a smart tool-filtering system for my AI agent to cut costs. Turns out I was silently nuking my KV cache on every request. Here's what Claude Code does instead.

03-08 · 4 min read · Agent · LLM · Engineering

The End of the Six-Figure Engineer?

Tech companies spent two decades selling Wall Street on the asset-light story. Now they're buying hundreds of billions in GPU infrastructure — and the math has to balance somewhere.

03-04 · 6 min read · AI · Career

The End of the AI Pipeline Era

Why the shift from hardcoded workflows to sandboxed agents isn't just an architectural upgrade — it's a fundamental change in how we think about building AI systems.

03-03 · 8 min read · Agent · AI · Engineering

AI Search Traffic Is Up 527%. Here's How to Get Cited.

Traditional SEO is losing ground as AI answers replace search clicks. This is a practical guide to GEO — making ChatGPT, Perplexity, and Claude actually cite your content.

02-27 · 8 min read · AI · Marketing

Why AI Is Making You More Exhausted, Not Less

A Harvard study tracked 200 tech workers for 8 months and found AI made them more burned out, not less. The root cause isn't AI itself — it's that AI rewrote the cost structure of work, but our decision-making hasn't caught up.

02-26 · 7 min read · AI · Productivity · Career

Why Your RAG Can't Find What You're Looking For

A brand manager couldn't find an article that was definitely in the database. The problem wasn't a bug — it was an economics problem hiding inside a technical one.

02-25 · 8 min read · RAG · AI · Engineering

LLM Training Isn't Alchemy. It's a Three-Stage Rocket.

Most people think training a large language model means feeding it data, tuning hyperparameters, and hoping for the best. The reality is an entirely different class of engineering — and the stage everyone obsesses over matters the least.

02-22 · 6 min read · LLM · AI