Blog
Notes on machine learning by Mikhail Balkhaev.
-
Company knowledge: searching documents
Index the company wiki, write a full-text search MCP server, and teach the agent to search documents the way a person does.
-
MCP: connecting internal systems
Move the agent's tools into separate MCP servers, write our own server for the company database, and teach the agent to connect to any server from a config.
-
The agent loop and tools
Give the model tools for files and shell commands, and write the harness that turns a chat into an agent.
-
Talking to a model
Take a commercial model API and write a streaming terminal chat with history in 40 lines.
-
The foundation: a server, a database and deploy on push
Rent a VPS, set it up, write the first Bun service with Postgres and migrations, package it with Docker Compose and Caddy, and deploy it with GitHub Actions.
-
Concept: an agent that knows the company
What agent we are building, why a chat bot is not enough, and what parts the system is made of.
-
Concept: which automations pay off
How to find the processes where a business loses money, work out whether automating them is worth it, and what the system we build them on will look like.