Construction project director (14+ years, general contracting) who transitioned into AI systems engineering by automating his own domain. I build autonomous multi-agent systems that run in production for months without human intervention — not demos, but live systems with real users and real business load.
2 orchestrators, 8 agent profiles across 2 specialization tiers, geographically distributed (senior on a 24/7 VPS with 5 profiles, junior on a separate local machine with 3 more). Running autonomously for months with 29+ scheduled jobs. Coding agents (Codex, Grok Build, Claude Code) are tools the orchestrators drive — not agents.
Quality is enforced, not hoped for: Mixture-of-Agents verification (maker ≠ checker, different model families, REJECT by default), rules in code rather than prompts, 8-layer governance audit, failure taxonomy (REASONING / TOOL / MEMORY / ORCHESTRATION / EVALUATION / COST) with per-class recovery paths. Cost engineering: model routing by task class + zero-downtime model migration → ~9× LLM spend reduction.
Architecture, gates, and full docs: github.com/Gromykoss/hermes-agent-farm
A long-horizon agent on a live 2,700 m² construction site: parses foremen's raw chat messages, photos, and voice notes into GOST-standard construction reports (4,266+ messages, 2,786 photos processed into 14 PostgreSQL tables, 837 work-order codes). Daily report generation went from 1 hour of human work to 30 seconds. Built from 14 years of first-hand domain expertise — the reports it generates are the exact documents its author managed as a project director.
Telegram memes → token detection → enrichment (DexScreener) → AI analysis → MoA verification → signal. Conservative two-model agreement with REJECT default.
Matrix/Synapse, custom web client + Android app. v107 in production, censorship-resistant architecture.
Learning pace: from first agent (Jun 2026) to four production systems (Aug 2026) — ~3 months, coding agents as the primary toolchain.
The farm is a general-contractor mental model applied to software: specialized crews, independent inspection, no unchecked handoffs. I treat token economics as a first-class design constraint (prefix-cache-friendly prompts, cheap models for routine work, expensive reasoning only on failure signals). Quality control lives in code, not in hopes that a prompt will be followed.
Long-horizon agents, multi-agent coordination, agent evaluation and safety. Full-time, part-time, or contract — remote worldwide.
Write on Telegram Email me GitHub X (Twitter)