⌖ Location Moscow, Russian Federation
⌗ Experience 2+ yrs commercial AI engineering · 5 yrs academic ML/neural networks
◎ Education NUST MISIS — MSc IT & Predictive Analytics (2024) · BSc Automated Systems (2021)
⟳ Focus Production RAG systems · Multi-agent AI orchestration · LLM infrastructure
Languages
Frontend
Backend & Databases
DevOps & Cloud
AI / ML
Claude API · Gemini API · Voyage AI · Qdrant · ChromaDB · Ollama · LiteLLM · MCP
🔹 jev-factkeep-compaction — Claude Code plugin: Jev-guided context compaction that never erases a tool call. On blind held-out rounds it kept 304 of 336 preregistered facts, against 36 for the original (50 of 50 on the latest round). Maintained in deadczarvc labs
🔹 hermes-jev-compaction — the same rules as a context engine for Hermes Agent: v0.90.0 ports the factkeep engine (reproducible reads shrink to a note, observations keep their errors, ids and counts) and pluggable decider chains (TypeSafe Jev, or any registry chain such as OpenAI gpt-6-luna)
🔹 cxtml — cXTML, an XTML dialect for multi-agent prompts on any model.
🔹 Why facts are lost — where the original compactor loses facts, with proofs and per-call data
🇷🇺 Russian 
🇬🇧 English 
"In the beginning was the Word, and the Word was — INDEX. Without an index, a session is the Library of Babel. With an index — O(log n) navigation across 1M context."
Made with ReadmeForge · github.com/deadczarvc