Andrii Suruhov·1d agoProfile a slow Python data pipeline (cProfile → py-spy → snakeviz)This is for engineers with a Python data job that got slow. The usual response is to guess: rewrite the pandas part, add workers, cache…
Andrii Suruhov·Oct 12How to Build a Real RAG System (Not a Demo): Ingestion, Retrieval, EvalsConfident, well-cited, wrong answers? Usually a retrieval miss. A build order, working code, and how to tell which stage broke.
Andrii Suruhov·Oct 1mypy vs ty vs Pyrefly: speed is settled, so stop shopping on itThis is for teams running mypy today and deciding whether one of the Rust checkers replaces it. Two are worth choosing between: Pyrefly…
Andrii Suruhov·Jul 22uv: replace pip, venv, and pip-tools in your Python data projectThis is for engineers who run Python data and ML projects and are tired of stitching together pip, venv, pip-tools, and pyenv. A typical…A response icon1A response icon1
Andrii Suruhov·Jul 14Datoon: convert your data to TOON only when it actually helpsJSON is verbose, so swap it for something tighter and pocket the savings. That’s the move everyone reaches for. It’s also, half the time…A response icon1A response icon1
Andrii Suruhov·Jul 14Lacon: a token-lean data interface for agents (not another DuckDB-over-MCP)Most people think the bottleneck when an agent touches data is reasoning. It isn’t. The bottleneck is that the agent reads the whole file…
Andrii Suruhov·Jul 10Release management for data and ML — the discipline nobody ports from softwareEveryone thinks shipping a model is harder than shipping software. It isn’t. It’s the same problem with the safety rails sawn off.
Andrii Suruhov·Jul 6Organon: a lifecycle and relationship layer for your filesWatch an AI agent touch a codebase. It runs grep, opens a few files, guesses what's related, and hopes nothing else breaks. It starts every…
Andrii Suruhov·Jun 22The Dracula Effect in AIThe exhaustion doesn’t come from working more. It comes from never working less.
Andrii Suruhov·Jun 17Human-in-the-loop before SQL execution: why and howThe agent doesn’t fail when the SQL is wrong. It fails when the SQL is plausible.
