01
System structure, component boundaries, and reliability
AI Systems Architecture
Architecting system boundaries, model orchestration, and reliability guarantees for complex production AI systems.
Architectural patterns and R&D studies in AI Lab
AI Systems Architect & Consultant
Valentyn Vitkovskyi
I design, evaluate, and build complex AI systems — agentic workflows, knowledge/RAG systems, and AI-native products — with a focus on architecture, reliability, and real-world constraints.
Hands-on R&D platform and engineering research for artificial intelligence systems without hype or premature complexity.
Expertise
Areas of Focus
Four key architectural pillars addressing system structure, reliability, and real-world constraints.
02
Autonomous workflows, orchestration, and state management
AI Agents & Agentic Workflows
Architecting autonomous agent workflows, tool execution boundaries, failure recovery, and verifiable human-in-the-loop controls.
Agent failure and stress-testing protocols
03
Retrieval architecture, knowledge provenance, and verification
RAG & Knowledge Systems
Designing enterprise retrieval-augmented generation (RAG), verified source attribution, hybrid search, and retrieval evaluation.
Retrieval accuracy and benchmark experiments
04
Prototype to production, unit economics, and staged rollout
AI Product Engineering
Engineering AI-native products, feasibility validation, latency/cost trade-offs, and staged production rollout.
Production benchmarks and cost optimization
AI Lab
Engineering Studies
Public engineering evidence and breakdowns of real systems and validated architectures.
When Green Tests Lie: Coding Agents and Real Runtime Verification
Problem:
An investigation of a reproduced production failure in vitkovskyi.com: despite extensive automated test suites and green agent implementation reports, the enhanced JavaScript form submitted FormData (multipart/form-data), which the PHP endpoint deliberately rejected with HTTP 415. Mocked Playwright tests failed to detect the defect. The targeted engineering repair converted the payload to URLSearchParams, aligned CSP for Cloudflare Turnstile, reconciled cross-runtime field limits, and introduced real browser-to-PHP integration tests with zero endpoint mocks.
Perfume Engine: An Evidence-Backed Recommendation Prototype Without a Runtime LLM
Problem:
Perfume Engine began as a broad fragrance-intelligence concept, but was deliberately narrowed during an early iteration into a controlled validation prototype: 12 perfume identities, 6–8 editorial mappings, exact search, manual evidence collection, and deterministic recommendation logic. The experiment was not designed to inject an LLM at every layer, but to verify whether a recommendation system could remain explainable, reproducible, and provenance-first. Because Perfume Engine subsequently evolved into a broader system, this study documents a specific historical architecture experiment rather than the full current product architecture.
Research
Recent Thinking
Analytical breakdowns, comparative evaluations, and architectural frameworks for informed decisions.
When an AI Product Does Not Need an LLM: Deterministic Inference, Provenance and Controlled Uncertainty
An engineering decision framework for choosing between generative LLMs and deterministic inference in production runtime. Why replacing an explicit knowledge base with stochastic prompts erodes reproducibility and trust, how to structure claim-level provenance, and which objective criteria determine where a foundation model creates genuine value versus unnecessary cost and liability.
Large PDF Translation Is a Workflow System, Not a Translation API
An architectural investigation into translating voluminous, structurally complex PDF documents. Why the naive abstraction of "PDF in → translation API → translated PDF out" breaks down across multi-hundred-page enterprise documents, how to structure an intermediate representation (DocumentIR), and why pipeline resumability, checkpoints, and layout reconstruction matter far more than model choice.
Simple Before Distributed: When an AI System Actually Needs Workflow Orchestration
An architectural complexity framework for AI workloads: identifying when synchronous execution suffices, and when queues, distributed workers, and durable workflows are genuinely required based on a 12-factor evaluation matrix.
Principles
Working Principles
Engineering and architectural principles that govern our approach to complex AI systems.
A system’s real constraints are defined by structure, boundaries, and data flows, not by the latest library or provider.
Assumptions belong in test harnesses, failure logs, and benchmarks before they belong in architecture decisions.
The best architecture is the simplest one that solves the failure modes reliably. Complexity is earned, not assumed.
Cost per query, latency percentiles, data privacy, and failure tolerance determine viability more than benchmark scores.
Where failure consequences are non-trivial, agent autonomy yields to explicit verification and human-in-the-loop controls.
About
Valentyn Vitkovskyi
Valentyn Vitkovskyi is an AI Systems Architect & Consultant focused on complex AI systems, agentic workflows, RAG/knowledge systems, and AI-native products. His approach centers on assumption testing, explicit trade-offs, first-party evidence, and designing under real production constraints. On vitkovskyi.com, he publishes research, architectural breakdowns, and AI Lab materials.
Work With Me
Discuss a Project or Architecture Review
If you are designing a complex AI system, planning agentic workflow deployments, or require an independent architectural review — submit a structured overview of your requirements for asynchronous review.