artemlegotin.com • LinkedIn • Telegram • email

I’ve been a software engineer since 2013, building AI products since 2020.

I focus on agentic LLM systems, generative AI, and recommender systems.

Currently, I’m Head of AI at exactly.ai.

Previously I co-founded neural.love and verstka.io.

• Outcome-aware context allocation: testing whether a learned controller can choose how much evidence a RAG system reads, with answer and citation checks.
• AffineFlow: learning how a robot responds over time, with a small control-cost gain in simulation.
• FamilyTiles: 20% less weight storage for two related language models, with exact reconstruction and measured inference limits.
• ExactKV: lossless LLM cache compression, showing when recovery costs more memory than compression saves.
• Dependency-aware verification: an experiment in verifying AI-written code without counting the same evidence twice.
• Relational Belief Compilation: keeping relationships between uncertain facts so decisions can show when the evidence isn’t enough.
• Customer Pilot Evidence Pack: turns search-pilot results into auditable reports, tracking failed requests and checking whether the evidence supports a comparison.

• Pythurion – Helps coding agents write Python you’ll want to maintain.
• minesweeper – A seatbelt for agentic coding in untrusted repo context.
• agent-badge – A README badge for repos that ship with AI.

• move-merit – Math behind a beautiful move.
• chess-gaze – Turns a chess stream video into gaze estimates and a local 3D viewer.
• ChessML – Advanced chess analysis, board & piece detection, and other tools.