Роман

Roman Kirchenko
Business & Systems Analysis · AI Delivery · Knowledge Architecture
Remote, Ukraine (Kyiv time) · [відкрити контакти](див. вище в блоці «контактна інформація») · stratoatlas.com · [відкрити контакти](див. вище в блоці «контактна інформація»)
Available from November 2026 · B2B contract via Ukrainian sole proprietorship (FOP)

P RO F I L E

Business-oriented systems designer with three years of commercial experience in complex B2B/ERP software and hands-on
experience building an independent AI-assisted research and knowledge platform.
My work spans business requirements, process modelling, functional specifications, user interviews, data relationships, AI
integrations, validation workflows and knowledge governance. I combine product and systems thinking with practical
implementation using Claude, ChatGPT, Supabase/PostgreSQL, Cloudflare Workers and the OpenAI API.
My current focus is turning complex business and information processes into reliable, documented and testable operational
systems.

C O R E C O M P E T E N C I ES

Business & systems Discovery, requirements, business rules, process models, roles, states, transitions, exceptions, functional
analysis specifications
AI-assisted development LLM workflows, structured outputs, AI-assisted implementation and debugging, human validation,
operational monitoring
Knowledge architecture Taxonomy, metadata, entity relationships, knowledge lifecycles, source provenance, revision history,
integrity checks
Delivery collaboration Cross-functional specifications, decision matrices, scenario validation, user interviews, backlog and
priorities
Technology Supabase/PostgreSQL, SQL (basic), Cloudflare Workers, OpenAI API, Miro, Figma, Jira, Google Workspace
Languages Ukrainian (native) · English C1 (two completed advanced-level courses)

P RO F ES S I O N A L E X P E R I E N C E

Founder · Knowledge Systems Architect & Researcher February 2026 – Present
StratoAtlas · independent R&D project · Remote
• Founded and developed an independent research platform investigating structural failures and contradictions in complex
systems.
• Design and maintain a typed knowledge graph (322 typed relations of 13 kinds) linking research cases, structural
mechanisms and publications, with metadata, source provenance, controlled knowledge lifecycles and revision history.
• Built AI-assisted research workflows with Claude and ChatGPT on Cloudflare Workers, Supabase/PostgreSQL and the OpenAI
API, including an LLM evaluation workflow with structured JSON output, queue-based retry and run telemetry.
• Established human review, decision records and validation procedures that separate preliminary assessments from accepted
research conclusions.
• Introduced a four-point integrity check before public content and a separate resolution path for incomplete or inconsistent
knowledge records.
• Maintain an operational research backlog, priorities and development decisions.
• Output as of October 2026: 195 documented cases, 26 mechanism records, 20 published structural maps and 28 publications
with DOI.
Technical case study on pages 3 to 4 · Knowledge architecture: stratoatlas.com/knowledge-architecture

Roman Kirchenko · CV and technical case study 1/4

Senior Product Designer · Business Processes & Systems Logic March 2023 – March 2026
B2B ERP / trading platform · Switzerland-based client (NDA) · Remote
• Worked with the Product Manager and engineers to turn business requirements into implementation-ready specifications for
a complex ERP/CRM platform: decision rules, roles, constraints, edge cases and user-facing behaviour.
• Authored detailed system specifications covering entity hierarchies, state transitions, approval rules, dependent operations,
notifications and exceptional scenarios.
• Modelled workflows, business entities and state diagrams, connecting functional documentation with implementation-ready
Figma prototypes.
• Designed processes involving multi-level cargo verification, external participants, contractual changes and role-based data
access.
• Conducted user interviews and incorporated the feedback into product and interaction decisions.
• Worked in Jira-based development sprints, including planning and retrospectives.

Senior Product Designer January 2023 – November 2024
Integrio Systems · Reconizer / R-MOR · Canada · Remote
• Redesigned the navigation, information hierarchy and interaction structure of an enterprise cybersecurity platform.
• Worked with complex information models and operational interfaces used by commercial and public-sector organisations.

Independent Consultant · Product & Information Design April 2010 – Present
Remote
• Product design, information architecture, analytical interfaces and digital solutions for Ukrainian and international clients.

Project Manager June 2012 – February 2017
Happy Child Foundation · Ukraine
• Coordinated international humanitarian and medical projects involving multiple stakeholders, procurement, documentation
and donor reporting.
• Delivered three consecutive projects under the Japanese Embassy's Kusanone programme, all passing independent audit.
• Managed external partnerships and operational coordination across participating organisations.

E D U C AT I O N

PhD studies, Management 2026 – 2030
Zaporizhzhia National University · Research focus: management of interactions between enterprise business processes through structural
diagnostics

Master's degree, Management of Foreign Economic Activity 2024 – 2025
Zaporizhzhia National University

Bachelor's degree, Economic Cybernetics 2002 – 2006
Zaporizhzhia State Engineering Academy

Roman Kirchenko · CV and technical case study 2/4

StratoAtlas
AI-Assisted Research & Knowledge Pipeline
Supplement to CV · technical case study in AI-assisted workflows, human validation and knowledge governance · October 2026

StratoAtlas is an independent R&D project that studies recurring structural failures and contradictions in complex systems,
including AI, enterprise software and defence technology.
I designed and built its research and knowledge infrastructure using Claude and ChatGPT as AI-assisted development tools,
while retaining responsibility for architecture, requirements, validation and operational decisions. The stack is Cloudflare
Workers, Supabase/PostgreSQL and the OpenAI API.
The central problem: turn a continuous stream of fragmented information into structured, verifiable knowledge, without
letting an automated assessment or an unsupported interpretation become an accepted conclusion.
A · DISCOVERY CONTOUR (HACKER NEWS WORKFLOW, IMPLEMENTED)

LLM evaluation Candidate pool
Source Rule prefilter Queue Human selection
relevance, not truth temporary

B · RESEARCH CONTOUR (CASE PROCESSING, RECORDED STATE)

Validation Artifact Knowledge Publication
Intake
alternative mechanisms text, diagram, review graph, audit manual approval

↻ A failed check reopens validation. External review can reopen a published claim.
automated recorded state human decision

1 LLM-assisted discovery
• A scheduled worker identifies potentially relevant material. A rule-based prefilter selects what is worth evaluating, which
reduces unnecessary model calls.
• A queue consumer sends the selected material to gpt-4o-mini. The model returns structured JSON: material type,
research potential, evidence strength, possible structural mechanisms and missing information.
• Candidates above the configured relevance threshold enter a temporary pool marked reviewed: false. A person decides
whether further research is warranted.
• The evaluator uses an eight-second API timeout and queue-based retry.

The model evaluates research potential, not factual truth. A high score is not evidence that the proposed explanation is
correct. That is the job of contour B.

The wider discovery infrastructure has logged more than 2,500 worker runs across six workers since May 2026. Each run
records what it processed, what it saved and what failed. Workers use different ingestion paths, so these counters describe
activity, not stages of one funnel.

IMPLEMENTATION EVIDENCE · HN EVALUATOR (SIMPLIFIED OUTPUT SCHEMA AND CONTROL FLOW FROM DEPLOYED CODE)

MODEL OUTPUT STRUCTURE (SCHEMATIC) CONTROL FLOW (EVALUATOR V0.5, CONDENSED)
{ const MIN_CASE_POTENTIAL = 0.45;
"material_type": "incident | workaround |
contradiction | governance_gap | // queue consumer, per message
metric_inversion | reject", if (await candidateExists(item.id)) return duplicate;
"case_potential": 0.0, context = await fetchHNComments(item.native_id);
"evidence_strength": "weak|medium|strong", result = await classifyMaterial(item, context);
"production_context": true, // OpenAI call, AbortController 8000 ms
"observed_pressure": "one sentence | null", if (result.case_potential < MIN_CASE_POTENTIAL)
"missing_piece": "none | mechanism | ...", return below_threshold;
"suggested_followup_query": "... | null", await upsertCandidate({
"possible_mechanisms": [ ... ], hn_id, title, url,
"reason": "one sentence" cdsa_score: result.case_potential,
} cdsa_reason: encodeReason(item, result),
reviewed: false // human review next
});
// success: message.ack() / error: message.retry()

The model classifies relevance only. Every accepted candidate stays unreviewed until a person decides.

Roman Kirchenko · CV and technical case study 3/4

2 A conclusion that failed validation
Once a topic is selected, it enters a separate process of evidence collection, hypothesis, alternative-mechanism testing,
review and knowledge governance. The current research protocol records processing states, step outputs and explicit
permissions for publication and closure; earlier cases may follow different record structures. Publication is never the
automatic result of finishing a text.
A real case processed on 29–30 July 2026:

Event Recorded result

Initial diagnosis A mechanism and a causal sequence were documented.

Alternative- The assumed chronology could not be established.
mechanism review

Reopening Two mechanism checks returned the explanation for further examination.

Restructuring The case was divided into two episodes with separate checks.

Final mechanism A revised interpretation was recorded.
verdict

Editorial review Verdict “revise”: the draft overstated the causal account.

Correction and Required edits applied; publication approved the following morning.
approval

After publication A prospective reversibility test was added to the record.

22 recorded steps, about 26 hours from first processing to publication approval. The validation changed the initial
explanation and stopped an insufficiently supported interpretation from passing unchanged. Reviewer notes and identifiers
are omitted here.

3 An architectural decision: integrity before public content
During an integrity review of mechanism records and their case associations, I found two inconsistencies that could have
entered the public registry. A confirmed case lacked its required link to a mechanism. Another case had been attached to a
mechanism by keyword match rather than on the evidence.
I introduced a four-point integrity check that must pass before public content is written, and a separate resolution path for
records whose definitions are unsettled, instead of finishing them inline during writing.

Missing text is a documentation problem. Missing definition is a research problem. Treating both as writing tasks risks turning
incomplete knowledge into apparently validated information.

4 Governance and operational boundaries
The typed knowledge graph (322 typed relations of 13 kinds) links cases, mechanisms and publications; sources, decisions
and revisions remain traceable.
Candidate records and permanent research records serve different purposes. Candidates are cleared from the pool. In the
inspected case, source URLs persisted in the permanent research records, while the original LLM evaluation details could
not be recovered from them. Preserving a minimal evaluation record when a candidate is promoted is an identified
improvement. Traceability of a source and traceability of an automated assessment are separate requirements.

AI accelerates discovery and analysis. Human review decides what deserves investigation, what can be concluded and what must be reopened.
The aim is not to remove judgment, but to make the process around it structured, repeatable and inspectable.
Public knowledge architecture: stratoatlas.com/knowledge-architecture

Roman Kirchenko · CV and technical case study 4/4