From Fine-Tune to NVFP4 on a DGX Spark

Published
Author

Every fine-tune of my Minecraft NPC is served in NVFP4, because the box is limited by memory bandwidth, not compute. How the pipeline from fine-tune to served checkpoint works, and Clef as the first build I published.

Read more

The Next Customer Is an Agent

Published
Author

Claude Code left out almost everything an IDE builds for a human, and that is why it works. What software looks like when the customer is an agent and the human is only watching.

Read more

Steal the Shape, Not the Mechanism

Published
Author

A transformer paper, a Minecraft NPC that forgets where it put the crafting table, and what happened when I borrowed the paper's shape one level up.

Read more

Rapid Product Development: RAD Without the Requirements Phase

Published
Author

Requirements gathering was insurance against expensive rework. Rework is cheap now, so I stopped buying it and let the requirements arrive as the decisions get smaller.

Read more

Iron Pickaxe Crafted

Published
Author

How two steps back and a new base model actually made progress.

Read more

Giving an LLM Eyes: a Low-Res Snapshot of the World

Published
Author

Spatial perception for language models is an open problem. My answer for Emil is a 48×35 text matrix, ray-marched from the world every turn — with a live view of the real thing.

Read more

I Rejected a Model at 44 tok/s. The Model Was Doing 108.

Published
Author

Swapping the base model under my Minecraft NPC: three configuration mistakes worth 2.5x, a container that crash-looped 23 GiB nine times, and a context-window constant denominated in a unit that changes size when the model does.

Read more

One Two Three Four Five Fine-Tunes That Never Learned to Think

Published
Author

Three checkpoints died because the chat template emptied every think block. The fourth, built to fix exactly that, learned to write 'Next: mine_block.' instead of thinking. The fifth was the fourth again.

Read more

A Perpetual Agent Must Never Wedge

Published
Author

Why 'context doesn't fit' has to be recoverable state, and the six-stage ladder in my agent runtime that guarantees it.

Read more

Optimistic Tool Calling: Acting Before the Model Stops Talking

Published
Author

Five papers, one design, one honest A/B: what speculative tool execution buys a 30B model on desk-sized hardware — and what it doesn't. Yet.

Read more

My NPC Tried to Place a Furnace It Didn't Have — 69 Times

Published
Author

Debugging a fine-tuned agent with an agent in the loop: a wrong hypothesis, a fake 100%, and a train/serve mismatch found in the timestamps.

Read more

Meet Emil: an LLM That Lives on My Minecraft Server

Published
Author

No scripted policies, no behavior trees. A 30B model, a Rust body, a memory, and a data flywheel on one desk-sized machine.

Read more

Agentic Development Doesn't Make You a Worse Engineer. It Removes the Tax That Experience Placed on Your Momentum

Published
Author

There's a trap that some developers fall into that nobody warns you about.

Read more

Don’t Call It Agile

Published
Author

How We Found Productivity without any Buzzwords

Read more

Next-auth integration

Published
Author

Implementing next-auth in a nextjs project using an external api with jwt token authentication

Read more

Astro.js and Tailwind

Published
Author

Project Showcase - Building this here with Astro.js and Tailwind

Read more