Oleksandr

[open contact info](look above in the "contact info" section) Portfolio:
Bratislava, Slovakia Oleksandr Vahabov github.com/
[open contact info](look above in the "contact info" section) AI / ML Engineer [open contact info](look above in the "contact info" section)

AI/ML engineer with hands-on experience in fine-tuning and quantizing LLMs, building RAG pipelines and multi-agent systems, and
shipping them as production backend services. Databricks-certified data engineer who also builds the surrounding platform: ETL
pipelines, CI/CD, and self-hosted infrastructure. Bridges research and engineering, and works directly with data analysts, data
engineers, and clients.
SKILLS
AI / ML Python, PyTorch, LLM fine-tuning, RAG, AI Agents (LangGraph, LandSmith), Vector databases.
Data & Backend Databricks, PySpark, Delta Lake, ETL, SQL, NoSQL, FastAPI, MongoDB, Redis.

Cloud & DevOps Azure, Terraform, Docker, GitHub Actions (CI/CD), Linux.
Languages English (C1), Slovak (C1), German (B2), Ukrainian & Russian (Fluent).

EXPERIENCE
Data Developer 03/2026 — PRESENT
Adastra Bratislava, Slovakia
• Led development of an internal migration engine and built Python automation scripts that generate mapping templates for
data analysts, removing repetitive manual setup from migration projects.
• Designed and maintained ETL pipelines and transformation workflows that deliver clean, reliable datasets for reporting and
decision-making.
• Designed and deployed a self-hosted internal platform on company servers, with all services containerized with Docker and
managed through Arcane
• Built CI/CD workflows in GitHub Actions for AI-assisted data transformation, assembling prompts dynamically from pipeline
metadata and routing them to an LLM through an internal integration tool.
• Introduced AI tooling into the team’s workflow and advised colleagues on AI use cases, solution design, and best practices.

AI Engineer 06/2026 — PRESENT
Leadgram
• Joined the engineering team of an AI product that finds and ranks people from public Telegram messages using hybrid semantic
search and a Rust backend.
• Worked on the Rust search service (hybrid retrieval, rank fusion, reranking, query understanding) and its parity tests against the
Python data pipeline.
• Contributed to the AI support agent (LangGraph, Azure OpenAI, RAG over a Qdrant knowledge base, safety guardrails, human
handoff).
• Operate in a production environment with Kubernetes, GitOps (ArgoCD), and Terraform-managed infrastructure.

AI Engineer 05/2025 — 02/2026
TeamChallange
PII Removal System
• Built a PII anonymization pipeline from fine-tuned XLM-RoBERTa NER models and rule-based heuristics, served through a
FastAPI API and web interface for real-time document redaction.
• Engineered batch ingestion from Azure Blob Storage with OCR; added OpenAI API as a secondary validator for complex
unstructured text; containerized with Docker.
• Result: 97%+ NER F1 and throughput of 500 documents/minute.
Client Fetcher (multi-agent lead system)
• Architected a LangGraph multi-agent system: one agent classifies incoming Telegram messages and stores leads, a second
retrieves leads and autonomously conducts the dialogue.
• Integrated it with async PostgreSQL and the Telegram API (Telethon) in a containerized Docker pipeline.
• Result: Autonomous lead detection, qualification, and escalation that replaced manual lawyer outreach.
Backend-Shop (microservices)
• Built four Spring Boot services (API Gateway, Auth, Product, Notification) with JWT, role-based access, Google OAuth2, RabbitMQ
messaging, and Docker deployment.

PROJECTS
Efficient Legal AI: Quantized LLMs with Fine-Tuning Bachelor Thesis, FIIT STU (2026/27)
• Built a Python evaluation framework (PyTorch, Transformers, bitsandbytes, llama.cpp) to measure how post-training
quantization (GGUF q4/q8, int8, NF4) affects Qwen2.5-3B across 145 LegalBench tasks (84.7k examples), using KV-cache prefill
reuse, dynamic batching with OOM recovery, and a verification gate against a reference scorer.
• Implemented QLoRA/LoRA training pipelines with checkpoint selection, 3-seed replication, and merge-and-requantize, plus a
generation-based scoring mode compatible with the original LegalBench protocol.
• Used rigorous statistics (balanced accuracy, task-stratified paired bootstrap CIs, McNemar) and found and corrected several
evaluation pitfalls.
• Results: quantization cost only 0.3–0.6 pp, while in-distribution fine-tuning gained +14.6 pp (69.4% → 84.0%), about 26x the
quantization loss. Merge-and-requantize kept 97–101% of the gain with higher inference throughput (21.0k vs 13.7k tokens/s).
LabelMe Object Recognition System Computer Vision
• Built a multi-label classification pipeline on LabelMe-12-50k, handling extreme class imbalance with data augmentation and a
ResNet-based architecture with skip connections.
• Result: 91% F1 in multi-class tasks and 80.2% subset accuracy in multi-label detection.
CERTIfiCATIONS
Sep, 2026 Databricks Certified Data Engineer Associate (credential)
Aug, 2026 IELTS Academic (C1)

EDUCATION
2024 – 2027 Bachelor of Science: Computer Science, Slovak University of Technology (FIIT), Bratislava, SK.

• Ranked in the top 7% of students; recipient of the faculty merit scholarship.