You're signed in. Welcome to kevinsomany.com.
Applied projects that put each lesson to work — live instruments you can operate.
Project
The agent is not hallucinating when it tells you to fix a bug you fixed yesterday. It is quoting the file contents it read at 2:00pm, faithfully, from a session that has no way of knowing the file moved.
View project →Project
A stronger prompt can make a review more thorough. Splitting the work decides whether the fourteenth file is read as carefully as the first. This project builds a multi-pass review pipeline from scratch and measures the difference.
View project →Project
Prompt instructions shape behaviour. Hooks decide what the application will allow. This project builds both hook points from scratch and shows why prevention has to sit before execution, not after it.
View project →Project
A better prompt can make a mistake rarer. A gate in your code can make a specific mistake impossible. This project shows where each belongs, and builds a refund tool that refuses to run until its prerequisites are met.
View project →Project
A good multi-agent system is not just the right subagents. It is a coordinator that can actually call them, and handoffs that carry every claim's origin with it. These are implementation problems, not theory problems.
View project →Project
Split a hard job across several AI agents and it can get faster or it can get chaotic. The difference is a coordinator that owns the plan, controls what each agent sees, and checks whether the combined result actually answers the request.
View project →Project
The gap between an AI that talks and an AI that does is a loop your program runs around the model. Here is that loop, built up one idea at a time and traced through a real support request, start to finish.
View project →