I’m Andrii. I build reliable AI systems, from complex data to real decisions.
00 / FEATURED RESEARCH
PUBLISHED COLLABORATIONBetter agents don’t just answer.
They learn when to escalate.
Co-author of support-agent reinforcement learning research at SoftServe, in collaboration with Weights & Biases and OpenPipe.
Relevant evidence supports a grounded resolution.
MY CONTRIBUTION
Multidimensional LLM-as-judge evaluation, grounding checks, and escalation behavior.
Trap tickets and out-of-distribution cases that expose brittle agent decisions.
Training-pipeline defect diagnosis, tool-parameter fixes, and checkpoint evaluation.
01 / SELECTED ENGINEERING
Explore the systems I’ve helped build. Choose a sample and see how each decision works.
ANONYMIZED PROJECTS · ILLUSTRATIVE INTERACTIONSFrom complexity to clarity.
Documents become typed fields, checked rules, and a deliberate review decision.Complex documents become typed, validated decisions.
Critical actions stay behind explicit human review.
Find the signal.
A question searches a constellation of knowledge and converges on supporting evidence.Technical manuals become a grounded conversation.
Retrieval, reranking, and local models working together.
Make the invisible visible.
Two illustrative detector paths expose differences before sensitive spans are masked.A reusable evaluation framework for privacy models.
Versioned inputs, typed adapters, inspectable results.
Intelligence, under control.
A request passes through policy, reaches a specialized tool, and leaves a trace.Specialized agents with runtime guardrails and traces.
Language models orchestrate. Python does the math.
Have documents, knowledge or workflows that need a reliable AI system?
Discuss your project03 / THE PERSON BEHIND THE SYSTEMS
3 years in AI. 6 years in IT. I build software that makes intelligence useful.
I build across the stack: Python services, retrieval pipelines, agent workflows, and evaluation infrastructure. I’m most useful where a promising idea needs clear architecture and working software.
I also help engineering teams build that capability themselves — through architecture reviews, hands-on workshops, and reusable reference implementations.
04 / MY WORKING TOOLKIT
Python · Agent systems · Retrieval · Evaluation · Azure
Python · FastAPI · Pydantic · asyncio · PostgreSQL · Docker
Azure AI Foundry · LangChain · NeMo · RAG · Tool calling
MLflow · Langfuse · pytest · LLM-as-judge · OpenTelemetry
Azure ML · Managed Identity · APIM · Azure Functions · Teams
02 / HOW I THINK
The real engineering is everything around it.
My work lives at the intersection of applied AI,
reliable software, and human judgment.
Need AI that holds up in production?
Typed contracts and deterministic rules own validation, calculations, and critical state changes.
Evaluate grounding, tool use, and failure behavior. Make every experiment reproducible.
Identity, network boundaries, observability, and integration belong in the architecture from day one.
Bound the agent’s autonomy. Give uncertainty a deliberate path to human review.