Skip to content

Serhii Zabolotnii — AI Researcher / Engineer / Solution Architect

Reproducible AI systems for research and high-stakes domains

AI researcher (Dr.Sc.) building NLP and LLM systems for regulated domains. I design evaluation and evidence tracing so that AI results can be checked, not just trusted.

AutoSci Research Framework TRACE

AutoSci Research Framework

Public

Author · research-system architect

A reproducible research framework built on top of the open AutoSci research agent. It carries work from idea selection and literature review through experiment design to manuscript preparation and submission tracking, with a chain of evidence from each number in a manuscript to the artifact that produced it.

  • Research agents
  • Reproducibility
  • Evidence chains
  • Multi-model orchestration

TRACE

Public

Author · framework architect

An engineering framework for trustworthy agentic AI in operationally critical domains: a four-layer reference architecture, 17 measurable trust metrics and the Computational Parsimony Ratio. Follow-up work defines runtime assurance contracts for high-risk AI agents.

  • Trustworthy AI
  • Trust metrics
  • Runtime assurance
  • Agent architecture

Capability map

Three areas of work. Select a project to jump to its card.

Capability mapResearch systems: AutoSci Research Framework. Trustworthy AI: TRACE, TRACE AI Gate. Applied AI case studies: AI module for the Supreme Court of Ukraine, ResearchIQ, LiveMedCard, Clinical AI in a regulated domain.AutoSci Research FrameworkResearch systemsTRACE AI GateTRACETrustworthy AIClinical AI in a regulated domainLiveMedCardResearchIQAI module for the Supreme Court of UkraineApplied AI case studies
Research systems
Agentic systems that carry research from idea to submission with a verifiable chain of evidence.
Trustworthy AI
Frameworks and metrics that make agentic AI measurable and auditable in critical domains.
Applied AI case studies
Production and research work in regulated domains, shown within its disclosure limits.

Selected work

  1. AutoSci Research Framework Author · research-system architect Research agents · Reproducibility · Evidence chains · Multi-model orchestration
  2. TRACE Author · framework architect Trustworthy AI · Trust metrics · Runtime assurance · Agent architecture
  3. TRACE AI Gate

    Public

    Author · engineer

    A deterministic runtime-assurance gate that stands in front of an AI model or agent. Each recommendation is checked against limits taken from the operator's own documents, pinned by hash, and gets one of three verdicts: pass, violation or undetermined, so a missing sensor or rule never counts as a pass. The live demo replays 7,311 open drilling records from Utah FORGE well 58-32.

    • Runtime assurance
    • Deterministic guardrails
    • Evidence trail
    • Drilling data
  4. AI module for the Supreme Court of Ukraine

    Public case study

    Feasibility R&D lead · co-architect · engineer

    For the modernization of the Supreme Court's Database of Legal Positions, led R&D on LLM correctness over real court decisions, co-designed the AI-module architecture and implemented legal search over Ukrainian text. The R&D informed the go/no-go decision for launch. Customer: Expertise France.

    • Legal NLP
    • LLM evaluation
    • Legal search
    • Ukrainian language
  5. ResearchIQ

    Public

    Author · product architect

    A multi-tenant research-analytics application that aggregates faculty publications from ORCID and OpenAlex for academic reporting. Deterministic entity resolution is combined with an LLM agent layer.

    • Entity resolution
    • LLM agents
    • Research analytics
  6. LiveMedCard

    Public

    Author · engineer · hackathon MVP

    A living medical record for an older patient, built as a hackathon MVP. It merges discharge summaries and lab results into one timeline, flags drug interactions, excessive doses and worsening trends with deterministic rules, and answers questions with a pointer to the source document. A language model may only propose a fact; it is accepted only when found verbatim in the document.

    • Deterministic safety rules
    • Evidence grounding
    • Health records
    • Human-in-the-loop routing
  7. Clinical AI in a regulated domain

    Under NDA

    AI researcher · NLP data analyst

    At a US healthtech company, led evaluation methodology and annotation practice for clinical information extraction, supported the transition to LLM-based extraction, and coordinated medical QA, Python development and labeling work.

    Client work under NDA: no system names, architecture, data or metrics are shown.

    • Clinical NLP
    • Evaluation design
    • Annotation methodology

Evidence