Нікіта

Nikita Voloshyn
Software Engineer | [відкрити контакти](див. вище в блоці «контактна інформація») | [відкрити контакти](див. вище в блоці «контактна інформація») | github.com/nikita-voloshyn |

Poznań, Poland (remote)

SUMMARY

Software engineer who ships production systems end-to-end — LLM agents (LangGraph, MCP), backend and APIs (FastAPI,
Django/DRF), data pipelines, and the infrastructure under them. Built the agentic layer of a social-commerce platform, including an
eval harness that fails CI when a prompt or tool change breaks tool selection. Shipped a webcam-based vision product solo across
applied ML, backend, desktop and infra, and built distributed crawling/ETL platforms for multiple clients. I take work from scope
through design and shipping, then measure whether it worked.

EDUCATION

BSc Computer Science — Adam Mickiewicz University in Poznań in progress

CORE COMPETENCIES

Python (async) FastAPI Django / DRF PostgreSQL & SQL REST APIs ETL & Data Pipelines LLM Agents & MCP

LangGraph RAG & Hybrid Retrieval Applied ML / Computer Vision Rust TypeScript / React Docker & CI/CD

EXPERIENCE

Social-Commerce Platform (name under NDA) Jan 2026 – Present

AI Engineer
Poznań, Poland · Part-time
Built the platform’s agentic layer end-to-end — an LLM moderation agent orchestrated over a Model Context Protocol toolset — owning
agent design, evaluation and the serving platform behind it. Models served through AWS Bedrock as a multi-provider access layer,
routing across model families on cost and quality behind one provider-agnostic interface.
Orchestrated the agent as a stateful LangGraph workflow (retrieval → policy gating → decision → action) with checkpointed state and a
full audit trail; ambiguous cases escalate to a human instead of being guessed, so the failure mode is a review queue, not a wrong
decision.
Built an eval harness that replays real runs and diffs the tool-call trajectory against a golden one — a prompt or tool change that breaks
tool selection fails CI instead of shipping. Pydantic contracts on every tool boundary turn malformed calls into loud, retryable failures
rather than silent data corruption.
Designed an MCP server (ASGI/FastAPI) exposing the catalog as a typed toolset, split across two servers: an admin surface behind
constant-time bearer auth, and a public read-only connector exposing only PII-free data, with no write tools and no path to the admin
ones. Underneath: FastAPI + Pydantic REST API, JWT and Google/Apple OAuth, PostgreSQL + SQLAlchemy/Alembic, async
Celery/Redis workers, Docker Compose, Nginx/TLS, Cloudflare, GitHub Actions CI.

Moru Tech Feb 2026 – Present

Software Engineer
Poznań, Poland · Self-employed
Sole technical owner of an eye-wellness product shipped end-to-end — a vision-training web app plus a passive desktop widget —
covering applied ML, backend, desktop, data and infrastructure.
Applied ML: real-time gaze estimation on MediaPipe in two runtimes (browser/WASM and a Python/OpenCV desktop sidecar) at 20–30
fps. Per-user calibration trained and served in-browser with TensorFlow.js, refit online from quality-gated samples with automatic rollback
— tripled the share of time estimated gaze landed on the intended target (22% → 64%). Reproduced a vague user complaint as head-
pose drift and shipped detection plus an affine re-fit: drift at a changed sitting position fell from 29% to 11%. Privacy by design
throughout: raw camera frames never leave the device — only derived signals reach the server, and a CI check fails the build if that ever
changes.
Backend & data: Django 5 + DRF REST API from scratch — auth and identity, subscriptions, consent-gated metric ingestion, admin
surface; self-hosted GoTrue with JWT validation across magic-link, password and Google OAuth. PostgreSQL with Row-Level Security
and versioned migrations, and a restore that is rehearsed rather than assumed: wiped the full data directory and recovered end-to-end.
Rest of the stack: React 19 / TypeScript SPA and PWA with an offline write queue and 12 locales (Vitest suite grown from 65 tests to
1,011); Rust / Tauri 2 desktop widget with signed auto-updates on macOS and Windows; one VPS on Docker Compose, Caddy (TLS),
Cloudflare, GitHub Actions.

Stealth Startup Sep 2025 – Feb 2026

Backend Engineer / Data Engineer
Poznań, Poland
Owned a distributed catalog-ingestion platform end-to-end. FastAPI task-enqueue API with URL canonicalization and automatic root-
category detection, feeding a 4-stage pipeline (crawler → paginator → reader → progress reporter) wired through durable Redis queues.
Separate processing flags and timestamps per record, so an interrupted run resumes instead of restarting and never double-writes.
Modeled the category graph in PostgreSQL (root → leaf → group → product) on SQLAlchemy 2.0, evolved through Alembic migrations
with backfills on live data. Extraction via Playwright plus direct collection from the internal GraphQL API; exports to JSONL/CSV/XLSX
with resumable image sync.

Freelance / Self-Employed Aug 2024 – Sep 2025

Python Developer
Odessa, Ukraine · Independent contractor
Reverse-engineered hidden and internal APIs for e-commerce and market-research clients: analyzed network traffic (browser dev
tools, HAR captures) to find undocumented GraphQL/REST endpoints, replicated their auth and session handling, and collected data
directly — bypassing page rendering entirely.
Async collection at client scale: asyncio with aiohttp/httpx for concurrent requests and asyncpg for writes, plus proxy rotation and rate
limiting tuned per client. Owned the delivery side too: scoped requirements with non-technical clients and agreed the output schema up
front.

E-commerce Data Company (name under NDA) Feb 2024 – Jul 2024

Backend Engineer / Data Automation
Odessa, Ukraine
Built Python/Selenium scrapers and Django backend services collecting product data from online marketplaces; added structured
logging and process monitoring.

PROJECTS

Regulens — Agentic RAG over EU Regulation open source

Agentic RAG over 8 EU digital regulations (AI Act, GDPR, DSA, DMA, Data Act, Data Governance Act, CRA, NIS2) with article-level citations,
self-verification, and model routing by query complexity. Hybrid retrieval (dense embeddings on pgvector/HNSW + BM25 + deterministic
citation lookup) over 3,880 structurally-addressed chunks from EUR-Lex/CELLAR. Cross-encoder reranking and structural sub-chunking
raised context recall@10 from 0.76 to 0.95 and citation-accuracy F1 from 0.56 to 0.71 — beating full-context stuffing of a frontier long-
context model (F1 0.67) at ~93× lower cost per query. Orchestrated in LangGraph with a self-checking groundedness node and a Postgres
checkpointer; cost-aware routing cut average generation cost by 44% with no loss in citation accuracy (F1 0.71 → 0.73); hand-verified
golden set published as an open HF Dataset; CI fails on metric regression.
Python · LangGraph · PostgreSQL + pgvector · sentence-transformers + cross-encoder reranking · bm25s · pytest · github.com/nikita-voloshyn/regulens ·
huggingface.co/datasets/wenitwa/regulens-eu-citation-qa

NanoVec open source

Exact KNN vector database in Rust, written from scratch with no external indexing libraries, queried by an agent directly over MCP.
Rust · MCP · candle · embeddings · github.com/nikita-voloshyn/nanovec

SKILLS

Languages: Python · TypeScript / JavaScript · Rust · SQL

Backend: FastAPI · Django / DRF · asyncio · aiohttp / httpx · Pydantic · Celery · pytest · REST APIs · JWT / OAuth · PostgreSQL (RLS) ·
SQLAlchemy 2.0 · Alembic · Redis

AI / LLM: Anthropic Claude API · AWS Bedrock (multi-provider model access, cross-family routing) · LangGraph · Model Context Protocol ·
agentic patterns (tool-calling loops, escalation, fallbacks) · structured outputs · prompt engineering · custom eval harnesses · RAG (chunking,
hybrid dense + BM25, cross-encoder reranking, grounded citations) · embeddings · vector search (pgvector)

Data: ETL and event-driven pipelines · Redis queues · idempotent, resumable jobs · schema migrations and backfills · Playwright ·
BeautifulSoup · GraphQL/REST reverse engineering · proxy rotation / anti-bot handling
ML / CV: MediaPipe · OpenCV · TensorFlow.js · real-time gaze estimation · feature engineering · online learning · Kalman filtering

Infra: Docker · Docker Compose · GitHub Actions CI/CD · Nginx · Caddy (TLS) · Cloudflare · self-hosted VPS · least-privilege service design ·
structured logging

Frontend / Desktop: React 19 · TypeScript · PWA (offline-first) · Tailwind CSS · Vitest · Rust / Tauri 2 Spoken: Ukrainian (native) · Russian
(native) · Polish (C1) · English (B2+)

CERTIFICATIONS

Model Context Protocol: Advanced & Introduction — Anthropic, 2026 · Data Engineering Specialization — DeepLearning.AI, 2026 · Machine
Learning Specialization — DeepLearning.AI / Stanford, 2024 · Certified Professional — Neo4j, 2026