Едуард

Eduard Chirva
AI / LLM / ML Engineer
LLM integration · RAG & agents · applied ML / computer vision · Python backend

[открыть контакты](см. выше в блоке «контактная информация») · [открыть контакты](см. выше в блоке «контактная информация») · t.me/qvartzgolden20 · github.com/qvartzggolden01 · [открыть контакты](см. выше в блоке «контактная информация»)
едуард-чирва 517018411 · Ukraine · Remote · Ukrainian (native) · Russian (fluent) · English B1

SUMMARY

Engineer with about two years building applied-AI and LLM systems hands-on: RAG pipelines, LLM agents and
integrations, self-hosted model serving, and computer vision, on a Python backend. I design these systems end to
end — architecture, retrieval and prompt/guardrail strategy, deployment — and I write code with AI-assisted tooling
while deepening my core fundamentals as I go.
I come to AI from IT: about six years growing from system administrator to Head of IT at УВК Нексус (team of up
to 30, a 200-person company, infrastructure built from scratch). So I read a system whole — from infrastructure
and serving up to the API and the agents on top — not only from the code.
Open to AI / LLM / ML engineering roles, fully remote, any timezone.

EXPERIENCE

Python / Applied-AI Developer — internal tools & LLM integration
УВК Нексус 200 staff) · 2024 2026 · Ukraine · Remote

Moved from IT management into hands-on development, focused on LLM integrations, RAG, computer-vision
automation, and internal Python tools, delivered end to end and used in daily operations.
Owned the AI/ML and automation projects listed below — from architecture and retrieval/prompt design through
deployment and production support.

System Administrator → Senior System Administrator → Project Manager → Head of IT
УВК Нексус · 2020 2024 · Ukraine

Joined as the company's sole system administrator and built the IT function from scratch as the business
scaled from a small office to a 200-person group.
Grew through senior sysadmin and project manager into Head of IT: ran all core infrastructure (local networks,
servers, RDP, VPN, SIP telephony, surveillance, end-user support), introduced ITIL-style processes, and led a
team of up to 30.
Owned budget-to-productivity optimisation: kept infrastructure efficient while scaling, then cut RDP and
maintenance spend during a later cost-optimisation phase.

PROJECTS

AI / LLM / ML

Conversational LLM Platform Commercial · delivered

Conversational platform on LangChain / LangGraph with a Weaviate vector database SQLite fallback), automatic
language detection, voice transcription, and background workers for reporting and export. Built as an observable
LLM pipeline rather than one-off prompt scripts.
Stack: Python 3.12, FastAPI, LangChain / LangGraph, Weaviate, PostgreSQL, Docker

Self-Hosted LLM Serving + Autonomous Agent R&D · in use

Self-hosted LLM platform running open models locally on a single consumer 24 GB GPU, with three serving modes
(an always-on pool, a hot-swap manager, and a layer-split GPU RAM mode for large models). Custom FastAPI
gateway with token auth and a multi-stage request pipeline; hybrid RAG LanceDB BM25, Reciprocal Rank
Fusion). Fully air-gapped. Includes a small code agent Plan / Execute / Audit / Debug) with a stdlib-only web UI.
Stack: Python, FastAPI, llama.cpp CUDA , GGUF, LanceDB, BM25, asyncio

Computer-Vision Automation Pipeline Commercial · delivered

Applied-ML automation across three layers: a trained YOLO model (dataset prep, augmentation, transfer learning,
ONNX / TensorRT export), a weighted action orchestrator, and programmatic control of the target application. Runs
continuously across many parallel instances.
500+ concurrent device instances, 24/7
Stack: Python, YOLO, OpenCV, ONNX / TensorRT, LLM API

Async Multi-Source Data Engine Personal · running 24/7

Long-running async backend: concurrent loops under a supervisor pattern, a multi-stage data pipeline with
deduplication and multi-factor scoring, a fault-tolerant Redis wrapper, and cost-aware routing across three LLM
tiers.
8,500 lines of Python across 36 files, built for continuous fault-tolerant operation
Stack: Python, asyncio, SQLAlchemy 2.0 (async), PostgreSQL, Redis, Docker, Claude API

AI SYSTEMS — DESIGNED (SPEC COMPLETE)

Maritime-Risk RAG Platform M.A.R.E. XO Design · spec complete

Architecture for a grounded, source-attributable question-answering system over a controlled corpus: hybrid
retrieval (dense + BM25, fused with Reciprocal Rank Fusion), an agentic query router with function calling and
text-to-query, chunk-level citations, guardrails, composite confidence and human-in-the-loop review, plus an
evaluation set. Full technical specification produced (available on request).
Stack: Python patterns → Node/TS · Vercel AI SDK, MongoDB Atlas Vector Search, BM25, RRF, LLM APIs

Local Voice + LLM Assistant MFO Design · spec complete

End-to-end architecture for a fully local, real-time voice assistant under an on-prem constraint (call data can't
leave the perimeter). Pipeline: telephony Kamailio / FreeSWITCH → VAD + semantic turn-taking → STT → local
LLM serving (SGLang, continuous batching, prefix caching) with constrained decoding and PII redaction → TTS,
grounded on hybrid RAG BGE M3 LanceDB) with tool calling for live data. Complete implementation strategy
produced (available on request).
Stack: SGLang, Qwen3, Pipecat, Kamailio / FreeSWITCH, Parakeet, Kokoro, Silero VAD, XGrammar, Presidio, LanceDB

BACKEND & AUTOMATION

Lead-Processing & Automation Platform Commercial · delivered

Replaced a manual spreadsheet-and-chat workflow with a single platform. Node.js + Python over a simple JSON
lines protocol, a mobile-first multi-role UI, and a messaging-integration layer with rate-limit and reconnection
handling.
150+ active users, 130K+ records in daily production use; runs on 2 CPU / 4 GB RAM
Stack: Node.js, Express, Python, PostgreSQL, Docker, supervisord, WebSocket, OpenAI API

Async Request Classifier Public repo · code sample

Async multi-source classifier: semaphore-based rate limiting, exponential backoff, a provider abstraction via ABC,
and Pydantic schemas. Public on GitHub as a readable Python sample I can walk through.
Repo: github.com/qvartzggolden01/netpeak-request-classifier
Stack: Python, asyncio, Pydantic, ABC

TECHNICAL SKILLS

AI / LLM 2 yrs LLM integration OpenAI, Claude APIs; open self-hosted models), agentic systems
(autonomous agents, LangChain / LangGraph), RAG BGE M3 BM25 RRF hybrid
retrieval, LanceDB, Weaviate), prompt engineering, structured output, guardrails, PII
redaction

LLM serving / self-hosting 1 2 yrs SGLang, llama.cpp CUDA ; quantization trade-offs GGUF, FP8 , KV-cache sizing &
capacity planning, continuous batching, air-gapped deployment

ML / Computer Vision 1 yr YOLO (training, transfer learning), OpenCV, ONNX / TensorRT

Backend 2 yrs FastAPI, asyncio, aiohttp, httpx, Pydantic v2, SQLAlchemy 2.0 (async), REST APIs,
background workers, rate-limit handling

Languages 2 yrs Python (primary), SQL, JavaScript / Node.js; some TypeScript

Data & Infrastructure 3+ yrs PostgreSQL, Redis, SQLite, LanceDB, Weaviate; Docker / Compose, Git, Linux, basic
CI/CD

Networking / SysAdmin 6 yrs MikroTik, local networks, VPN, RDP, SIP telephony, server management

How I work Design-first (problem → cost → success metrics → then the technology); AI-assisted
implementation while reading and owning the code; deepening core fundamentals;
end-to-end delivery

EDUCATION

National Technical University "Kharkiv Polytechnic Institute" NTU "KhPI") 2023 2026

Bachelor's degree — Computer Engineering (specialty 123

College "Osvita", Open International University of Human Development "Ukraine" 2016 2019

Junior Specialist (associate-level) — Law

Kremenchuk Higher Vocational School No. 7 2013 2016

Vocational diploma — Information Processing & Software Operator