Today I work at Workwatch, a contractor at a petrochemical plant, automating production tasks in code.
Turns a Telegram export into markdown you can hand to an AI. No network, no keys: everything stays on your computer.
Where people are idle, and for how long
Wearable-device and BLE-zone data
Site managers at a large retailer could finally see where and how long people were idle. Delivered to a pilot.
⏽ beads-hud http://127.0.0.1:7777 │ bd 20 готово · 2 в работе · 5 ждут │ ctx 43% │ 5ч 38% 1ч19м │ нед 8% │ $34.34 │ Opus 5·1M xhigh
Project tasks and documents where you already look two hundred times a session.
# Task dispatcher
Classify every request with the classifying-tasks skill before acting.
Heavy tasks → delivering-heavy-features skill. Shipping → shipping-a-release skill (explicit ask only).
Short rules that catch the mistakes a coding agent repeats on every project.
Four years in applied AI. Before that, a large Russian retailer of over a hundred thousand people and an agricultural company of fifteen hundred: using data to improve how cleaning, delivery and the stores themselves worked.
I look for value in the low-level processes nobody examines. Decisions go into PRD and ADR documents, so six months later the team is not arguing the same point again.
The data is almost always already there: devices keep writing, the database keeps growing. I bring it down to one screen a manager can act on.
I take apart the low-level processes that run on people and habit, and automate them in code.
I set up a working process with agents for developers: from framing the task to verification before release.