Pavel Volkov · Engineering Notes
Field notes on building verifiable autonomous AI systems: executable specifications (Agentic SDD), adversarial multi-model verification gates, and digital employees with shifts, charters, and defined risk boundaries.
I tested Jev, TypeSafe's decision model, on lead classification in my Derzhi Lida bot. On the very task Jev is built for, cheap general-purpose models were more accurate and no more expensive.
Read →Five parallel AI teams made me the bottleneck again. What changed in two and a half months: a lead with written authority, rules drawn from the statistics of my answers, a search over my decision corpus, and stall detection from outside.
Read →How a local Qwen 2.5 model controls five Jasminers, five climate zones, and water-pipe freeze protection while mining offsets the cost of useful heat.
Read →How we benchmarked 12 models to replace OpenAI Codex for plan reviews in Findrates, why Tencent Hunyuan 3 won, and how we engineered process safety in macOS.
Read →Why canonical SDD turns into dead documentation, and how we built a working Agentic SDD pipeline based on production signals, adversarial model reviews, and BDD checks.
Read →The first 18 days of an autonomous digital employee on a prediction market: fifty shifts without my involvement, eleven hypotheses, the first profitable month, and two limits it never crosses without the owner.
Read →Why dozens of AI sessions force you to delegate not just tasks but their management: a Telegram concierge, a memory of decisions, routing, and result tracking.
Read →How my Polymarket trading infrastructure became a proving ground for an autonomous digital employee with a charter, shifts, memory, and SRE on-call duty.
Read →How I separated roles, checks, and releases so that several AI teams can work in parallel without my constant supervision.
Read →