# Nazarii

[@nazar1ous](https://hacklab.so/nazar1ous)

Nazarii is Principal Machine Learning Engineer at Leaptree\.AI and Founder at Aizexis, based in Warsaw. No AI coding activity was synced in the last 30 days.

Principal ML engineer at Leaptree\.AI and Aizexis co\-founder\. Builds LLM fine\-tuning, agent evaluation and production MLOps systems\.

Location: Warsaw

Joined: 2026-07-22

0 followers · 0 following

## Links

- [GitHub](https://github.com/nazar1ous)
- [LinkedIn](https://linkedin.com/in/nazarii-kuspys-802a85213)

## AI usage (All time)

No activity in this period.

## Accomplishments

### [deblurGAN\_PRO](https://github.com/nazar1ous/deblurGAN_PRO)

Python repository from 2021\. Its README says the project aims to get rid of motion blur on photos and names PyTorch\.

### [ai\-assistant\-telegram\-bot](https://github.com/nazar1ous/ai-assistant-telegram-bot)

Python repository from 2024\. Its README describes a simple attempt at an AI\-assistant bot in Telegram, based on GraphRAG\.

### [twitter\_api\_usage\_for\_socialization](https://github.com/nazar1ous/twitter_api_usage_for_socialization)

Python repository from 2020, described as letting a user find friends located nearby so they can socialize by going out\.

### AMLD EPFL Edge AI conference speaking

Lists an Edge AI Track talk titled Efficient AI for Image Deblurring on Mobile Devices at AMLD EPFL in Lausanne\. Event date is unspecified\.

### PyTorch distributed\-training contributions

Profile reports three major PyTorch contributions and a 15% distributed\-training performance improvement; specific pull requests are not supplied\.

## Experience

### Principal Machine Learning Engineer — Leaptree\.AI

2025\-09 – Present

Warsaw, Poland

Architected automated fine\-tuning for Llama 4\-Scout and other MoE architectures, including dual\-head output for structured scores and natural\-language rationales\. Built a Numerical Token Loss calibration layer addressing LLM grading logit bias; the profile reports 20% better score alignment\. Designed event\-driven AWS SageMaker MLOps with ETL retraining and batch\-transform inference on multi\-terabyte data\. Used PEFT/QLoRA and GPU\-memory sharding for constrained training\. Established a question\-driven taxonomy of verifiable agent\-behavior metrics across Salesforce CRM domains\.

### Founder — Aizexis

2025\-06 – Present

Warsaw, Poland

Entrepreneurial sales/partner role delivering working, tailored AI systems for mid\-size companies\. Headline identifies the member as co\-founder\.

### Principal Machine Learning Engineer — Thumbprint Furniture

2024\-02 – 2025\-09

Los Angeles, United States

Built recommendation and real\-time computer\-vision systems\. The profile reports a recommendation service for 50,000\+ daily users at 99\.9% uptime, a 32% conversion increase and $280 higher average order value\. Created AWS ECS/Lambda/SQS inference processing 10,000\+ furniture images per hour below 200ms latency\. Autoscaling reportedly reduced compute costs 45%/$180,000 annually while supporting fivefold traffic spikes\. Built A/B testing across 15\+ metrics and one million\-plus interactions\.

### Technical Team Lead / Machine Learning Engineer — Paysera

2023\-02 – 2024\-02

Warsaw, Poland

Led a cross\-functional team of 12 \(seven ML, four backend, one frontend\) on an enterprise chatbot serving 500,000\+ monthly active users across eight countries\. Built Graph\-RAG using Pinecone/Weaviate; the profile reports 60% faster query responses and 94% answer accuracy\. Built fine\-tuning over 10 million\-plus customer interactions, with reported satisfaction improvement from 3\.2 to 4\.7/5\. Designed GDPR/PCI DSS compliance controls and availability infrastructure; model compression/caching reportedly reduced inference costs 55%/$400,000 annually\.

### Senior Machine Learning Engineer — SoftServe

2018\-07 – 2023\-02

Warsaw, Poland

Led an eight\-person 3D generative\-AI R&D team; the profile reports six client demos securing $2 million in project funding\. Built Stable Diffusion/post\-processing pipelines for Trimble 3D map assets and multi\-GPU avatar training \(reported 72\-hour to eight\-hour reduction on eight A100s\)\. Developed edge object detection with a local question\-answering agent for meal identification and manufacturing circuit\-board anomaly detection\. Built Jenkins/Docker/Kubernetes MLOps for 15\+ production models with automated testing, quantization and distillation\. Reports 80% less deployment time, 300% faster inference and management of an $800,000 annual AWS/GCP budget\.

### Junior ML Researcher — SoftServe

2017\-06 – 2018\-07

Ukraine

Junior ML research role\. The role description lists an AMLD EPFL Edge AI talk on efficient mobile image deblurring and three major PyTorch contributions supporting distributed\-training performance\.

## Education

### Bachelor of Applied Science — Computer Science — Ukrainian Catholic University

2019\-06 – 2023\-04

Computer Science BASc program, June 2019 through April 2023\.

### Master program — Information Technologies — Ukrainian National Forestry University

Computer Science / Information Technologies master program with an education period listed as 2023–December 2024\.