Особисті дані приховані

Kostiantyn Krykunov AI/ML Engineer · multi-agent orchestration · LLM infrastructure ·
Polars-first Kyiv, Ukraine · [відкрити контакти](див. вище в блоці «контактна інформація») · [відкрити контакти](див. вище в блоці «контактна інформація») ·
github.com/Kkrykunov · ORCID 0000-0001-8301-6021

==============================================================
EXPERIENCE
AI/ML Engineer (core development) · Miltech / J&Y Jul 2025 – present
Built an MCP server stack (14+ servers) with a Triad multi-agent pattern (planner /
executor / critic) and a CLAUDE-BRIDGE local↔cloud state-synchronization layer.
[lesson: one MCP server per semantic surface, no overlapping tools]
Cost-optimized model routing: routine subagent tasks delegated to cheaper models
(GitHub Copilot CLI + smaller local models), Claude reserved for high-value
reasoning — ~60% token savings.
LoRA fine-tuning (personal GPU + cloud) and Q4_K_M-quantized models under
Ollama for local inference of routine Claude Code subagent tasks.
Multi-agent CLI reverse engineering (26-day sprint): full reverse of a modern
Qualcomm chipset using Ghidra Headless Analyzer + Hexagon SDK wired in as tools
inside the agent loop (GUI stripped); reusable pattern library across the chipset
family.
Stock-modem reverse engineering: deblock of restricted bands + 3GPP power-class
change via RFNV manipulation and AT-handler patches; firmware dump + root
extraction on a commercial 5G dongle.
Prompt engineering as executable contracts: CLAUDE.md v2 + userPreferences v3
(trigger words, lens frameworks, epistemic markers [ε:high/mid/low/?]).

Data Analyst Consultant · World Health Organization (WHO), Ukraine Country Office Jul
2024 – Jun 2025
First production use of the Claude API in the HERAMS team: integrated LLM-based
standardization into the WHO Ukraine primary-healthcare data pipeline ahead of
country-office adoption.
Normalized 16,000+ PHC facility names (Python NLP + FuzzyWuzzy + Claude API) at
98% accuracy, delivered 1.5 months ahead of schedule; deployed as a REST API.
Built a VAR / Bayesian VAR / Granger-causality engine for an MPH thesis on mental-
health forecasting (Google Trends, multi-country panel); lag order selected by
AIC/BIC. Same kernel re-pointed to BTC prediction by swapping the input layer.

Medical Data Analyst (part-time, alongside clinical anesthesiology) Feb 2014 – Jul 2024
Clinical research & municipal-hospital analytics consulting
Exploratory data analysis and full analytics pipelines for state / municipal clinical
hospitals (consulting): data cleaning → EDA → modeling → reporting;
payment/coverage and service-availability benchmarking against NSZU standards.
Predictive clinical-risk model: post-cardiac-surgery dialysis-risk Random Forest,
ROC-AUC 0.78 (standard clinical risk scores ≈ 0.70–0.75; confirm exact baseline
before quoting).
Patient risk stratification: +25% over baseline.
Reproducible ML pipelines (Python / R) on REDCap-sourced multi-site clinical-trial
data; supported 7 PhD / doctoral research projects.

==============================================================
EDUCATION
MSc Computer Science (AI & Machine Learning track) · Neoversity Jun 2026 – 2028 MSc
Public Health (M&E focus) · NaUKMA (Kyiv-Mohyla Academy) Sep 2023 – Jun 2025 Thesis
methodology: VAR / Bayesian VAR / Granger causality on a multi-country panel. MD in
Medicine · Bogomolets National Medical University 2004 – 2010

==============================================================
SELECTED PROJECTS
NSZU clinical-coding integrity engine (regional cardiovascular-surgery hospital,
2026) — FLAGSHIP. Production-grade ML decision-support on 15,677 real electronic
medical records (38 features). Detects DRG/ДСГ misgrouping and B→A severity-
reclassification candidates from clinical features (ICD-10 + ACHI intervention codes,
regex-parsed from free text; AR-DRG-derived expected grouping). Three-model
ensemble (5-fold stratified CV, ROC-AUC): LDA 0.908, Logistic Regression L2 (class-
balanced) 0.882, Random Forest n=500 0.894; consensus voting flags cases for expert
review (decision-support, not autocoding). Built as a reproducible pipeline: compiled
Parquet source + schema, input validation + output reconciliation scripts, fixed
random_state, permutation-importance feature ranking, refactored shared
constants/utils. (No manual ground-truth labels — supervised on existing coding;
tariff weights reconstructed from secondary sources and flagged unverified.)
TaperCalc — PK/PD-driven medication-taper engine. Hyperbolic taper default; PMID-
first verification; UNVERIFIED flag for unsourced values. Covers SSRIs, SNRIs, TCAs,
BZDs, gabapentinoids, MAOIs, antipsychotics, mood stabilizers.
Cross-asset correlation (Apache Spark) — distributed lag / correlation analysis across
financial assets with confidence intervals (Python + Hadoop/HDFS).

==============================================================
SKILLS
AI/ML infrastructure: MCP server orchestration (14+ stack), Claude Code with subagents,
multi-agent architecture (Triad pattern), Ollama local + cloud, LoRA fine-tuning, cost-
optimized model routing, prompt engineering as executable contracts. Machine Learning:
ensemble methods (LDA / Logistic Regression / Random Forest), permutation importance,
class-imbalance handling, time-series (VAR / Bayesian VAR / ARIMA / auto-ARIMA),
Granger causality, regression, classification, A/B testing, model selection (AIC/BIC, ROC-
AUC), 5-fold stratified CV. LLM / RAG: LLM integration (Claude, GPT, Gemini, LLaMA);
RAG (LangChain, ChromaDB, embeddings via sentence-transformers, reranking, citation
tracking), RAGAS evaluation — prototype/roadmap level. Programming: Python
(advanced; Polars-first / NumPy / scikit-learn / PyTorch / FastAPI / asyncio), SQL
(PostgreSQL/MySQL), R, JavaScript, Solidity (read+modify), Bash, Kotlin Multiplatform,
C/Rust (read). Big Data: Apache Spark, Hadoop (full course), distributed processing.
Reverse engineering: Ghidra Headless, Hexagon SDK CLI, AT-command driver patching,
firmware dump/extract (binwalk, mtdutils), multimeter probing. Web3: Solidity smart
contracts (read+modify), blockchain transaction analysis, Web3 frontend integration.
Tools: Git/GitHub, Docker, AWS, Linux, Jupyter, VS Code, Claude Code, GPU programming.
Languages: Ukrainian (native), English (C1), Czech (B2)