DENIS SUDAKOV[ru/en]

Building infrastructure for AI agents, language model post-training,
and local AI systems.
Moscow, Russia

Email · GitHub · CV

~~~

ABOUT

I build practical AI systems, with a focus on core infrastructure for
agents and post-training workflows.

Most of my work focuses on reliable, composable systems for AI agents
and local-first agentic infrastructure. I believe the current ecosystem
lacks foundational tooling for building useful, extensible agents that
are easy to share, modify, and integrate. My goal is to design and
implement these missing primitives as open, composable systems.

Currently studying geology at Lomonosov Moscow State University while
spending most of my time working on AI systems and ML research.

~~~

PROJECTS

 - su-agent (WIP)
   Local agent runtime and backend for running extensible AI agents on
   a user's machine. Exposes a system API for multiple client apps,
   skill/tool packages, session and memory management. Built around
   modularity, multi-client support, and local execution.

 - su-memory (WIP)
   Extension for su-agent (also compatible with Pi coding agent).
   File-based persistent memory and skill incubator. No vector search,
   embeddings, or RAG — just files.

 - su-subagent (WIP)
   Extension for su-agent (also compatible with Pi coding agent).
   Minimal sub-agent delegation via spawned agent processes. No shared
   context, no nested conversations, no framework.

 - su-ASR
   OpenAI-compatible speech recognition server. Real-time streaming,
   keyword biasing, multilingual support, low-latency inference.

 - Madiz
   Radiation monitoring system on a stratospheric satellite platform
   (RPi 4 + custom sensors). Handled hardware integration, onboard
   processing, compression pipeline, and control logic.

Early-stage

 - su-code
   Coding agent multiplexer on top of su-agent, Codex-style interface.
   Originally forked from t3-code.

 - su-claw
   Personal extensible agent with memory and skill management.
   Inspired by Hermes-style agent architectures.

Research

 - GSPO-PyTorch
   Implementation of Group Sequence Policy Optimization — an RL
   algorithm for LMs addressing GRPO limitations via sequence-level
   importance sampling.

 - aiijc-701w
   RL fine-tuning of Qwen models for mathematical reasoning. Dataset
   construction, benchmarking, GRPO-based training pipelines.

~~~

WRITING

 - Coming soon — research notes and system design write-ups

~~~

INTERESTS

 - LLM post-training for agentic environments
 - Local AI infrastructure
 - AI developer tooling

~~~

TEACHING

 - Instructor, International Summer School
   MSU University Gymnasium (2026)
   Teaching AI, machine learning, and modern agent systems to
   high school students.

~~~

EDUCATION

 - Lomonosov Moscow State University
   BSc Geology (2025–2029)

Independent coursework

 - Stanford CS336 — Language Modeling from Scratch
 - Stanford CS229 — Machine Learning
 - Stanford CS230 — Deep Learning
 - Harvard CS50AI
 - Harvard CS50
 - Harvard CS50 Web