Articles by Grizzly Peak Software

I made Kodiak, my open-weights decision model, and Jev, TypeSafe AI's closed one, play Zork by choosing from the game's valid commands: same harness, same seeds. In about 20,000 moves, not one was invalid. Jev played better, then walked into a grue while its own danger answer said "yes" at 0.95. Here's what that, a no-model baseline and a 35x question-wording effect taught me about System 1 / System 2 cascades.

Kodiak is an open-weights decision model: give it a document and typed questions, and it answers all of them in one pass, in 8 ms, with calibrated confidence and an honest "I can't tell." I built it on my DGX Spark with Claude Code in three days. Here's how it works, what it costs, a 3 a.m. plot twist, and where it stands against bigger models.

A weekend building a guard for agent tool calls with Jev: rules set the floor, calibrated confidence handles judgement, and the Claude Code hook that ended up guarding the agent that built it.

Jev, TypeSafe's new decision model, is trained to be calibrated, not to please. Here's what that means, what it's good for, and how to use it from Claude Code without building an agent framework.

I started load testing our Next.js backend-for-frontend in Postman. I finished in k6. Here's why I switched mid-project, what Claude did for me along the way, and the one refactor that turned a pile of response times into numbers leadership actually asks for.

I got Eon Systems' whole-brain fruit fly emulation running on an NVIDIA DGX Spark: 138,639 neurons wired from a real fly's connectome. It reproduced the paper's sugar-to-feeding result, needed a CUDA 13 fix to build the faithful GPU path, and ended up driving a simulated body that walks and turns on its own descending-neuron signals.

Every KDP problem is cheapest to fix before you hit publish — and most account disasters were preventable in minutes. The pre-publish audit a 200+ title operation runs on every book: the instant-stop gate, the six risk domains, and why risk is two numbers, not one.

Written from a real enforcement file that includes two KDP account terminations — both reversed. How to triage which emergency you're in, what to do (and not do) in the first 48 hours, what appeals can and cannot achieve, and the post-restoration audit nobody tells you about.

You wrote it yourself and Amazon still flagged it as "similar to other books." Similarity enforcement isn't a plagiarism check — here are the six ways original books trip it, the keyword-field mistake that has terminated accounts, and the five-minute pre-publish checks that prevent all of it.

The "content quality" email is short and generic on purpose. Here's how to read it, which of the four enforcement modes you're actually in, what to do in the first 24 hours, and the response mistakes that turn a fixable flag into a terminated account.

This article reveals the concrete operational thresholds where Postgres becomes cost-prohibitive, backed by real query diagnostics and migration case studies that prove when to act.

I'll show why feature flags accumulate debt when not actively managed, then give the exact metrics and process to stop ignoring them. You'll build a sustainable flag strategy that prevents codebase bloat.

You'll learn why staging environments deceive your team about production readiness, and gain practical validation techniques to catch failures before deployment. By the end, you'll have a checklist to replace trust with proof.

Stop chasing uptime metrics that mask real failures. This article reveals three user-impact metrics that prevent outages, with code to instrument them without rewriting your stack.

I benchmarked four local models on an NVIDIA DGX Spark across three context depths and a purpose-built agent reliability harness. The fastest model wasn't the best one, and the model I'd been running for weeks was quietly fabricating data.

This article exposes how standard retry strategies exacerbate system crashes, not prevent them. You'll learn to replace naive retries with circuit-breaking resilience patterns that keep your system running during chaos.

On-call cycles aren't just tedious: they actively devalue engineering skills. This article exposes how they sabotage promotion paths while offering sustainable alternatives.

This article debunks microservices for small teams with real migration pain points and concrete alternative patterns. You'll walk away with a single-process architecture blueprint that scales better than microservices.

Exposes how CI caches lie about artifact integrity, not just cache misses. Teaches concrete validation patterns to prevent real-world build failures before they break production.

Most engineers overestimate savings until they run the numbers. The real threshold is when your inference volume exceeds 150 requests/day.