AI/ML engineer. I build agent systems and take them to production.
Before that — 25 years of architecture, production and engineering management.
Currently: an autonomous development pipeline for a desktop C++/Qt product. It takes a real user ticket, writes the code, proves the fix in the live compiled GUI, passes an independent audit on a separate machine, then commits and answers the user.
Measured over equal 8-week windows, before and after launch:
- accepted tickets per week: 4.7 → 61.0
- report to first fix, median: 2.8 → 0.8 days
- share delivered by agents: 73%
- cost: $500/month in subscriptions for all of it — the same load at list token prices would be ~$5k/month, or $12.11 per accepted task
Writing (in Russian):
Production code is under NDA. The architecture, the measurement method and the limitations are public in those two.
Next here: an evaluation harness for coding agents on open-source repositories — the gap I named in part two.