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.

• FamilyTiles: lossless weight compression for related language models, saving 20% storage on the tested pair; includes CPU and Metal decoders.
• Adaptive tree-neurons: grows neural networks with trainable branches inside selected neurons, preserving existing predictions when capacity is added and testing where growth helps.
• Counterexample-guided operator coalescing: tests LLM weight sharing across layers to reduce checkpoint size, using low-rank corrections and difficult inputs to choose which layers to merge.
• FAM-Jet: compresses event streams for later time-window queries, testing whether retaining event age improves answers within the same storage budget.
• Reasoning allocation: tests a learned scheduler for the SCIP optimization solver, choosing search branches and time budgets while counting scheduling overhead against the deadline.
• AffineFlow: combines learned robot dynamics with actuator physics to predict motion at different timesteps for planning; evaluated on a simulated arm.
• Outcome-aware context allocation: tests context budgets for question answering. An evidence-first prompt raised correct answers with exact citations from 27 to 103 on 512 benchmark questions.
• Customer Pilot Evidence Pack: turns search evaluations into offline reports that keep failed requests visible and block provider comparisons when the evidence is incomplete.
• ExactKV: a lossless codec for LLM attention caches, showing why reconstruction memory erased the tested storage saving.
• Dependency-aware verification: tests how coding agents should choose between tests, reviews and documentation checks without treating repeated evidence as independent confirmation.
• Relational Belief Compilation: a decision engine that preserves relationships between uncertain facts, acts when all retained interpretations agree, and shows conflicting cases when they don’t.

• Pythurion: an agent skill for writing small, readable Python changes, with focused checks for correctness and performance.
• minesweeper: an agent skill that checks untrusted repository text for prompt injection before treating it as instructions.
• agent-badge: adds a README badge showing recorded AI coding tokens or estimated cost, with usage data stored in the repository.

• move-merit: uses engine analysis and a model of likely moves to explore a chess move’s quality, alternatives and difficulty.
• chess-gaze: estimates where chess streamers look from video and visualizes the results in a local 3D viewer.
• ChessML: reads chess positions from images using board detection and piece recognition; includes pretrained models and tools for building datasets.