[Automation]
AI teammates you call by name in Basecamp.
Mention Jack Sparrow on a Basecamp card and a minute later he answers there: a number, a log line, or a plan that becomes a pull request.
Read the post
The AI Agency blog.
All posts
How thirteen repos share one AI context
Each repo carried its own context file, so the same question got a different answer depending on which folder you had open.
[Engineering]Three rejections, and the guide that would have prevented all of them
We spent six months getting an MCP server through OpenAI's review. It was rejected three times, and not once because the server did the wrong thing. Every step below is a published requirement, with what we got wrong at each one.
[Engineering]The screening is a conversation. The 113-question form fills itself
How a phone app that listens in Gujarati took the retyping out of a school health study in Gandhinagar district.
[Case study]How we built NOVA for LD Expert, then handed their team the controls
What it takes to hand a client an agent they can direct themselves, and the testing pipeline that makes that safe.
[Case study]Why AI pilots fail to reach production
The demo proves the model can do the clever part. Production depends on everything the demo left out.
[Engineering]Fine-tuning or retrieval (RAG)? How to choose
Fine-tuning changes how a model behaves and retrieval changes what it knows, so the real question is which problem you have.
[Architecture]From a Basecamp card to a pull request
A card moves, and a few minutes later there is a comment on it, or a pull request.
[Agents]Designing escalation paths for AI agents
Where an AI agent should stop and ask, and how it earns the right to ask less over time.
[Agents]Generating help articles from pull requests
The diff tells you what changed. It does not tell you what the screen now says.
[Engineering]AI document extraction that survives a bad scan
Demo accuracy collapses on real intake, and the fix lies mostly in image handling, arithmetic checks and fast human review.
[Data]Keeping AI model costs down in production
Model spend rarely spikes: it drifts, one reasonable change at a time, until someone finally reads the invoice.
[Operations]Testing non-deterministic systems like LLM apps
Why exact-match tests fail on AI output, and how to build a suite that is stable, honest and keeps learning.
[Testing]From spreadsheet process to AI-assisted service
Read the formulas, time the copying, and give AI the reading and drafting while code keeps the numbers exact.
[Automation]