# Ilya Fedotov

> Engineering lead in AI infrastructure and MLOps. Builds and runs GPU
> infrastructure for training and high-throughput AI inference. Head of
> Engineering at Singularity Compute and Head of MLOps at SingularityNET,
> in parallel with deputy head of the laboratory at NaInt.

Eight years in infrastructure, four of them in MLOps. Works on cloud AI
platforms, GPU infrastructure, production model inference and the teams that
run them, across public clouds and owned hardware. Based in Russia, works
remotely. Russian native, English advanced.

Contact: fedotow.ilja@gmail.com

## Pages

- [CV](https://fedotov.io/cv/): Full record — experience, stack, open source, publication, education.
- [CV in Russian](https://fedotov.io/ru/cv/): The same record in Russian.
- [Blog](https://fedotov.io/blog/): Writing on GPU infrastructure and inference. No posts yet.
- [Blog in Russian](https://fedotov.io/ru/blog/): The Russian edition of the blog.

## Research

- [Good-Enough LLM Obfuscation (GELO)](https://arxiv.org/abs/2603.05035): With Anatoly Belikov, March 2026, cs.CR and cs.LG. A lightweight protocol for privacy-preserving LLM inference on untrusted accelerators, using fresh per-batch invertible mixing instead of MPC or FHE.

## Open source

- [Harness-engineering](https://github.com/Eljaja/Harness-engineering): An agent-first blueprint for repositories where AI agents plan, edit, test and evaluate.
- [deepagents_multiagent](https://github.com/Eljaja/deepagents_multiagent): Multi-agent experiments.
- [BobaClaw](https://github.com/Eljaja/BobaClaw): A coding-agent client in Rust.
- [ha-nlu-router](https://github.com/Eljaja/ha-nlu-router): Local voice and text control for Home Assistant with no cloud LLM: wake word, Whisper, a small intent classifier over a closed label set, deterministic slot filling and a safe/confirm/block policy.
- [home-assistant-tool-calling](https://github.com/Eljaja/home-assistant-tool-calling): Datasets for Home Assistant function calling.
- [BibaVPN](https://github.com/Eljaja/BibaVPN): A DPI-resistant SOCKS5 and HTTP tunnel over TLS and WebSocket. Most-starred repository.
- [vllm-omni-sparse-attention](https://github.com/Eljaja/vllm-omni-sparse-attention): Sparse attention on top of vLLM.
- [central-llm-gateway](https://github.com/Eljaja/central-llm-gateway): A fully self-hosted LLM platform with authentication, quotas, metrics and tracing.
- [GitHub profile](https://github.com/Eljaja): Everything else.

## Elsewhere

- [LinkedIn](https://www.linkedin.com/in/ilya-fedotov-517ab6231/)
- [Singularity Compute](https://singularitycompute.com/)