Nautiloid Protocol


AI Systems. Built to ship, not to scope

Nautiloid Protocol is an independent AI engineering practice.
B2B contracts. US-billed (ACH / Wire). US-based.
No account managers. No offshore handoff. One senior engineer, end-to-end.

 

What gets shipped

  • LLM-powered products — inference backends, agentic systems, RAG pipelines. Cloud or self-hosted.

  • Production AI systems — full-stack ownership from architecture to deployment. Not a proof of concept.

  • Custom model work — fine-tuning, LoRA, synthetic data generation. When off-the-shelf models don’t fit the problem.

  • On-prem / air-gapped — full ML stack, no cloud dependency required. When it’s a hard requirement.

Stack: Python, Rust, vLLM, llama.cpp, ONNX, FastAPI
Protocols: REST, WebSocket, Protobuf, JSON-RPC
Observability: Otel, Grafana, Clickhouse
Infrastructure: bare-metal, Proxmox

 

The Operator

– Valerii Verezhynskyi AI Systems Engineer

Employee #1 at Refact.ai. Builds systems 0→1, not Jira tickets

Opinionated on how AI systems should be built. Owns the infrastructure, the code, and the decisions.

Ships full stacks. Has opinions on your architecture. Will tell you when your approach is wrong.

Selected engagements:

[x] Refact.ai Nov 2021 – Jan 2025

Employee #1. Built data pipeline over 80M code repos. Co-authored dataset for Refact-1.6B-fim (SOTA HumanEval 2022). Engineered enterprise LLM inference backend, RAG over AST + vector search, agentic capabilities (SWE-bench compatible). Rust LLM client.

[x] BeeSensible Jul 2025 – Dec 2025

Interim Head of AI, 0→1. Full air-gapped ML stack in <6 months: training, storage, deployment, synthetic data generation, streaming inference, fine-tuned BERTs + LoRAs across 70 production labels.

[x] Coxit Feb 2025 – Jun 2025

Lead AI Developer. NER model for PII detection (WASM-compilable). MCP server with RAG over technical documentation, LLM-as-a-Judge evaluation.

→ Full résumé

 

Got a Problem?

You have an AI problem. It’s either not built yet, not working in production, or built wrong.

If you need it owned and shipped — not workshopped — 30 minutes. No deck. We figure out if it’s a fit or we don’t waste each other’s time.

→ Schedule a discovery call