Grigorii Kirgizov

AI Research Engineer

Senior AI research engineer with 10+ years across science and industry. I build systems with the capacity to learn and self-improve — at the intersection of agentic AI, formal methods, and evolutionary computation.

Skills & Knowledge

AI & Reasoning
Agentic AI · LLM (LangGraph · LiteLLM · etc) · Deep Learning · Reinforcement Learning · AutoML · Evolutionary Algorithms
Formal Methods
SMT Solvers · Type Systems & Semantics · Language Design (DSLs / LLVM) · Compilers · Complexity Theory
Mathematics
Category Theory · General Algebra · Graph Theory
Languages
Python · Rust · C++ · Kotlin · Java · Haskell · Prolog
ML & Data
PyTorch · Transformers · NN architectures · NLP · NumPy · Pandas · MediaPipe · ONNX / wasm
Web & Product
React · Vite · RxJS · Capacitor (iOS) · Supabase · Vercel · Product & GTM
Infra & Craft
Docker · Kafka · OpenTelemetry · Playwright · CI/CD · Architecture · Linux · git

Teaching & public-lecture experience · academic English (C2) · creative & product design.

Experience

2024 — Present

Noeon Research

Reasoning-AI startup · Tokyo

noeon.ai

Senior Research Engineerresearch

High-quality R&D across the company's core projects — in close collaboration with researchers and engineers — at a startup building next-gen reasoning AI from category theory, logical solvers, and agentic methods.

  • Designed and built a bidirectional bridge between the in-house formal DSL and SMT-solver languages — including formal translation of a coroutine-like language — foundational for the system's reasoning.
  • Ran R&D on solver coverage and speed, logical-representation trade-offs, and a subsystem for complex Sudoku solving.
  • Authored research reports and algorithm design for an ARC-AGI pipeline; built data-generation and benchmarking for the subgraph-matching algorithm.

Feb 2022 — 2024

ITMO University

AI in Industry Lab · Saint Petersburg

@aimclub

AI Research Engineerresearch

Leading development of GOLEM — an AI framework for graph optimization with metaheuristic methods (genetic, swarm). Driving architecture, research direction, and a small research team.

  • Led a full architectural redesign of the lab's primary project and its integrations with other teams' tools.
  • Launched GOLEM, a new framework for graph optimization, and initiated a research direction on adaptive evolution with reinforcement learning.
  • Contributed to 3 research papers (in progress); supervised interns and MSc students; taught classes on architecture and graph optimization.

Dec 2020 — Oct 2021

Grounding

V.V. Dokuchaev Soil Museum · Saint Petersburg

groundingwith.space

Exhibition Curatorart

Co-curated an international exhibition of technological art dedicated to soils — 28 artists, 20 art objects and installations across 2 venues, 3000+ visitors, ~100 media publications, and a parallel educational program of 10+ events.

  • Official partner of Ars Electronica 2021.
  • Short-listed for the Innovation 2021 state art award (best educational project).
  • Short-listed for the Kuryokhin Award 2022 (best exhibition).

Nov 2018 — Jan 2021

JetBrains

MPS — Meta Programming System

jetbrains.com/mpsmps-coderules

Software Engineerengineering

R&D on Coderules, a Prolog-like meta-language for declaratively defining type systems inside MPS, the IDE for domain-specific languages.

  • Significantly improved the meta-language's syntax and semantics.
  • Implemented type checkers for Java and (partially) Haskell.
  • Built a working prototype of an incremental, language-independent type-checking engine.

Talk — “Coderules, a new typechecking engine”, MPS Users Conference, Amsterdam, Oct 2019.

Oct 2017 — May 2018

Saint Petersburg State University

Bachelor Thesis

paper · ISP RAS

Researcherresearch

Developed an embedded C-like DSL and a library for programming small-scale heterogeneous microcontroller systems, working with LLVM IR and dynamic code generation.

Jul — Aug 2017

JetBrains

Software Intern

NLP Engineer (Intern)engineering

Built an NLP framework for extracting structured information for marketing research — entity extraction, web ontologies, and RNNs.

Nov 2016 — May 2017

Raidix

Data Science Intern

paper · CEE-SECR

Data Scientist (Intern)research

Improved industrial SSD caching algorithms with predictive machine learning — a research article followed by product integration.

Education

2019 — 2021

MA in Arts, Art & Science (with Honours) · ITMO University · Saint Petersburg · art.itmo.ru

2014 — 2018

BSc in Computer Science (with Honours) · Saint Petersburg State University · Software Engineering · se.math.spbu.ru

Selected Projects

GOLEMNeurIPS poster · GECCO '24

Graph optimization & learning

A framework for graph optimization and learning by adaptive, hybrid evolutionary algorithms. Built around two ideas — universality of the approach and adaptability of the algorithm — and used by several teams at the ITMO AI Institute.

aimclub/GOLEM

MPS CoderulesJetBrains

Type systems for DSLs

A declarative, Prolog-like engine for defining the type systems of domain-specific languages in JetBrains MPS — including a prototype for incremental, language-independent type checking.

JetBrains/mps-coderules

Focus Zenfull product · live

Attention-aware timer · Web & iOS

The first productivity timer that's aware of you and adapts to you — it follows your focus, pulls you back from distractions, and surfaces insights (and, I'd argue, it's the most aesthetic timer out there). Built on a research-informed attention-tracking state machine with custom gaze-tracking and brush-drawing. Solo, end-to-end: idea, research, engineering, product, landing, and go-to-market.

focuszen.io

clarcresearch solver

Contract-learning for ARC-AGI

A neurosymbolic ARC-AGI solver that wraps Poetiq's record-setting refinement loop in a machine-checkable layer: a typed DSL paired with z3 contracts over an abstract grid domain, sound by construction. Candidate programs are refuted before they ever run — 99.4% refutation power, zero false refutations across 13,641 checks — and an LLM-free synthesizer grows its own DSL to solve tasks with no model call at all. Verified invariants cut the median iterations-to-solve from 8 to 3, at zero harm.

gkirgizov/poetiq-arc-agi-solver

Arbitrage Graphagentic R&D

Crypto market microstructure

An autonomous R&D program that mines crypto-market microstructure for tradeable edges: a real-time arbitrage-graph engine feeding an agentic research loop that tests falsifiable hypotheses against live cross-market data and graduates the survivors toward execution (testnet → live). The founding premise — risk-free negative cycles — proved productively wrong (all stale-quote artifacts), and the meta-finding is sharp: every retail-scale micro-edge is real but sub-fee. One survivor cleared the bar — the tokenized-gold basis, a market-neutral spread uncorrelated with BTC, gold, and USDT stress.

gkirgizov/arbitrage-graph

Distributed Intelligence Environment2022

DIE · artificial life

An artificial-life project reproducing the emergence of distributed intelligence under environmental pressures, using neural cellular automata grounded in Evolution, Reinforcement Learning, and Active Inference.

gkirgizov/die

Publications & Talks