Nektar Insights

Insights

Thinking on AI, data, and building things that work.

I Built My AI a Second Brain That Survives the Context Window

Every Claude Code session starts blank. So I built the layer that makes it not start blank — a five-layer memory of facts, scars, rules, and narratives, held in check by one law so it never eats the context window it was meant to preserve. An interactive walkthrough of the whole architecture.

It's the Back-to-Backs, Not the Travel

Two years ago I got the Utah Hockey Club badly wrong. Right idea about what wears a team down, wrong model. The rebuild, the Mammoth's 2026-27 projection, and what actually moves a game: rest, not miles. The follow-up to the Utah Hockey Club post-mortem.

Context Window Optimization Case Study

My AI coding sessions were dying after an hour. Two fixes — one architectural, one 15 minutes — cut my context overhead by 75%.

What 1,449 Utah Vanity Plates Look Like When You Let Sonnet Read Them

I gave Claude Sonnet 1,449 photos of vanity plates from my camera roll. One batch pass — OCR, interpretation, category, vehicle, state — total cost $4.50. The interactive census, the methodology, and what the model got right and wrong.

I Was Treating Claude Like a One-Person Shop. So I Ran an A/B Test.

My Claude bill kept climbing while Codex and Ollama sat unused. I asked the obvious question — does delegating actually save anything? Here is what the numbers said.

Rebuilding my Utah Hockey Club submission, 21 months later

Public post-mortem of my 2024 submission to the Utah Hockey Club Summer Analytics Challenge. Seven specific failures, a rebuilt methodology, and what 21 months of unseen data did to the numbers.

Vibe Coding Has a Pilot Problem

AI is the most capable co-pilot ever built. But a co-pilot without a pilot just crashes faster. Three modes of working with AI that actually produce results.

My AI Suggested a Technique from Cancer Research. It Was Exactly Right.

Averages hide the story. Kaplan-Meier survival curves show you when loans default — not just if. How a technique from oncology transforms portfolio analytics.

Drives Beat Drops at Every Level. 304K Shots Prove It.

I analyzed 304,000 pickleball shots and built an interactive dashboard in an afternoon. Here's what the data says — and what it proves about building with AI.

I Almost Destroyed My Production Server Because the AI Said It Was Safe

The monitoring showed CPU at zero. The AI said it was a ghost server. It was running five production apps. Here's what vibe coding gets wrong about infrastructure.

How 84 Secret Versions Cost Me More Than My Entire Cloud Run App

I set a $5/month GCP budget alert. When it finally fired, Secret Manager was 55% of the bill. Here's how I fixed it in 30 minutes.

Vibe Coding Has a Security Problem

Vibe coding compresses the time developers accumulate third-party trust and secrets. A look at what that means after the Vercel April 2026 incident.

Nektar Rival Alerts: MVP PRD

A timing-and-emotion product for instant rival-loss alerts, one-tap roast generation, and private sharing.

By the Numbers: 78 Days of AI Engineering

Activity heatmaps, token-volume charts, model mix, and ROI breakdowns — the interactive data companion to 78 days of building with AI.

The Practice: What 78 Days of AI Engineering Actually Taught Me

How daily AI sessions compound into production software — the ratio, the feedback loop, the difference between vibe coding and engineering discipline.

FYBY: For You, By You

The FUBU of software. Why the person closest to the problem is now the person building the solution.

Smoke Tests & Harnesses in Vibe Coding

The testing discipline that keeps AI-generated code from blowing up in production.