Notes

Can I Build an AI Agent That Doesn't Write Slop?

I tested prompt engineering and local fine-tuning to build an AI copy editor. Fine-tuning beats prompting for voice, but a modular pipeline beats fine-tuning for maintenance.

Fine-Tuning Was the Easy Part

Tuning a small model to generate minimal API field masks took an afternoon. The hard platform problem is distribution: getting that fix past a single adapter, across dozens of APIs, and into the foundation models developers choose next.

A Model Router Needs a Scoreboard

Routing tasks to cheaper capability tiers is an easy policy to write, but it remains a hypothesis until held-out runs prove quality holds and measure the real cost of retries, latency, and rescues.

Builder Platforms Grow by Owning the Agent Loop

Give coding agents a tested path into your platform, measure what works, and use the results to improve activation and retention.

Loop Engineering Coding Agent

A public operating contract with four role overlays, a structural check, and 17 specified scenarios. It is not a behavioral benchmark yet.

DevX Is a Growth Function

The growth loop starts when DevX owns repeated developer friction, ships the fix, distributes the better path, and measures whether behavior changed.

Curated context beats raw model knowledge for building maps ↗

I built maps with AI to test where current models fail. Here's what worked, what broke, and why curated context beats raw model knowledge.

Native GeoJSON in BigQuery means no more transformation pipelines ↗

Query massive map datasets immediately. Here's how to skip complex transformation pipelines and process map data directly using BigQuery's native GeoJSON support.