Location Remote · Open to relocation
Featured Projects
CAE Copilot – Engineering AI Copilot
May 2026 – Present
CAE Copilot · Full project details ·
Docs (evidence)
·
API (Swagger)
- Engineering AI copilot for technical documentation, calculations, and traceable answers — not solver automation; human-orchestrated CAE workflows.
- LangGraph router with shared agent state: intent classification → RAG (knowledge/setup/scripting) or Pydantic calculation tools (beam, FoS, bolt joint, sections).
- Document-grounded RAG on CAE docs corpus (9 manuals): Qdrant vector store, page-aware chunks, mandatory citations in structured reports.
- Reliability-first: deterministic Python tools, golden Q&A eval (14 cases), offline fallback baseline 100% intent / citation / tool accuracy.
- Production layout: separate FastAPI and Streamlit Docker images, multilingual-e5-base baked offline, Helm/k8s deployment.
- Quality gates: lint, typecheck, pytest; eval harness invokes full graph per golden case with rule-based metrics.
System Architecture
Streamlit UI → FastAPI → LangGraph Router → Qdrant RAG + citations → Engineering tools (Pydantic) → Kubernetes (Helm)
Key numbers (public)
- CAE docs corpus: 9 manuals (ingested)
- Maturity stage: Stage 2 (RAG + citations, LangGraph, tools, eval)
- Golden Q&A eval: 14 cases (100% offline intent / citation / tool)
- Deployment: FastAPI + Streamlit (Qdrant, Helm/k8s)
SoccerPredictAI – Production ML & MLOps Platform
May 2025 – May 2026
SoccerPredictAI · Full project details ·
Docs (evidence)
- Domain-agnostic proof of production ML/MLOps discipline: a complete lifecycle engineered as a system, not a collection of notebooks.
- Automated data ingestion for 950,000+ football matches across 26 years of history from 3 heterogeneous sources (Selenium/Selenoid, Chrome DevTools Protocol XHR interception, REST batch).
- End-to-end MLOps: DVC for data versioning, Great Expectations for strict data contracts, and MLflow for experiment tracking with nested Optuna trials.
- Leakage-safe experiment framework: walk-forward CV across 6 data fractions × 5 models (30 MLflow nested runs); XGBoost selected via CV logloss only — no holdout peeking
- Tuned XGBoost hyperparameters via Optuna (100 trials, CV-only objective); holdout evaluated exactly once in dedicated final_train stage, preventing selection bias
- Dual inference API: sync FastAPI endpoint (<500ms p95) + async batch precompute via Celery/RabbitMQ — decouples prediction freshness from request latency
- Deployed on Kubernetes + Helm; Prometheus + Grafana monitoring; DVC pipeline versioning (all stages hashed for reproducibility)
- End-to-end CI/CD: GitLab CI (ruff lint + pytest), SOPS + Age secrets, Great Expectations data quality gates at raw/features layers
- Observability: 11 custom Prometheus metrics (separating queue latency from pure inference latency), Evidently AI for continuous feature and prediction drift detection, and Grafana dashboards.
System Architecture
Scrapers (Selenium/CDP) → Airflow (K8s Executor) → PostgreSQL + MinIO → DVC + Great Expectations → MLflow + Optuna → FastAPI + Celery/RabbitMQ → Kubernetes (Helm) → Prometheus + Grafana + Evidently
Key numbers (public)
- Holdout log-loss: 1.006 (bookmaker benchmark ~0.97)
- Holdout ROC AUC: 0.643
- Calibration (ECE): 0.004
- Test suite: 560+ (passing tests)
- Infra cost: <€30/month (self-hosted VPS)
Key Achievements
- Built and deployed CAE Copilot: traceable AI copilot for engineering workflows (LangGraph routing, Qdrant RAG with mandatory citations, Pydantic calculation tools, golden Q&A eval); Stage 2 shipped on Kubernetes.
- Led simulation-supported development of the H820 harvester — awarded AGROSALON Golden Star 2024 — with R&D-to-production validation at <5% deviation between model predictions and field telemetry. (Rostselmash)
- Built and deployed SoccerPredictAI — domain-agnostic proof of end-to-end MLOps discipline: 950,000+ matches across 26 years; served via FastAPI + Celery on Kubernetes with <500ms p95 latency.
- Reduced manual data extraction from ~8h/day to <30min/day (95% workload reduction) by deploying an OCR + NLP pipeline processing MSDS documents in 12+ languages. (Chemwatch)
- Established CFD/DEM capability from zero at Gomselmash (2015): trained 3+ engineers, published calibration methodology in CADFEM Review (2020).
Professional Summary
Engineer with 14+ years across three connected competencies: simulation (CFD/DEM, HPC, digital twins), production machine learning (MLOps, Kubernetes serving), and applied AI (LLM/RAG agents). Led a simulation/HPC team from 2 to 10 engineers and delivered R&D-to-production validation at <5% deviation vs field telemetry; shipped production AI systems end-to-end — CAE Copilot (traceable LangGraph + RAG copilot for engineering) and SoccerPredictAI (full MLOps platform). Verification-first mindset in every domain: models are judged against reality with explicit acceptance criteria, leakage-safe evaluation, and calibration. Open to remote roles and relocation.
Competency Map
Simulation (CAE & HPC)
14+ years of CFD/DEM for machinery: digital twins, calibration, field validation at <5% deviation
vs telemetry. Led a simulation/HPC bureau (2→10 engineers); H820 harvester awarded AGROSALON Golden
Star 2024; methodology published in CADFEM Review.
CAE & HPC portfolio →
Production ML & MLOps
End-to-end ML systems: data contracts, leakage-safe evaluation, MLflow/DVC, Kubernetes serving,
observability. SoccerPredictAI — 950K+ matches, 560+ tests, <500ms p95; Chemwatch OCR/NLP
pipeline — 95% manual work reduction.
ML & MLOps portfolio →
Applied AI (LLM/RAG)
Traceable LLM systems: LangGraph agent routing, RAG with mandatory citations, deterministic tools,
golden evals. CAE Copilot — engineering copilot shipped on Kubernetes with a live demo, docs, and
Swagger API.
Applied AI projects →
These competencies combine into different roles — Simulation Lead, ML/MLOps Engineer, or the hybrid I find
most exciting: applying AI inside CAE workflows (copilots, surrogates, digital twins).
The shared thread is a verification-first mindset: models judged against reality with explicit acceptance
criteria, in every domain.
Research highlights
Peer-reviewed digital-twin methodology (CADFEM Review) and an MSc dissertation that anchors the simulation
research program — relevant to calibration, verification, and reliable ML systems.
Highlighted
Improving the efficiency of the grain unloading technological process in the unloading system of a self-propelled grain harvester
2018 D. N. Ivanov · Gomel State Technical University (GSTU) — MSc in Mechanical Engineering (defended Feb 2018) · Thesis
MSc dissertation synthesizing simulation-based research on grain unloading systems — the methodological foundation for later DEM/digital-twin work and peer-reviewed publications.
PDF available on request.
Highlighted
Building a digital twin of grain crop material flow
2020 D. N. Ivanov, D. V. Dzhasov, A. N. Vyrsky · CADFEM Review #7 (30) — peer-reviewed industry journal · p. 15 · Article
Flagship publication: digital twin of bulk material flow with DEM calibration against physical tests, test-rig methodology, and validation workflow — directly transferable to production ML (data contracts, calibration, verification vs reality).
Read article (PDF)
View all publications & proceedings →
Engineering Leadership & Program Ownership
I approach engineering systems as products: define acceptance criteria, build automation, operationalize verification, and make outcomes measurable.
My leadership experience comes from scaling simulation/HPC teams and workflows that support real production timelines.
- Built and led a simulation/HPC engineering group from 2 to 10 engineers; hiring, mentoring, onboarding, and technical direction.
- Delivered computational workflow automation for simulation engineering: standardized templates, Python tooling, and repeatable pipelines adopted across projects.
- Owned end-to-end delivery from R&D to production/field validation: stakeholder alignment, risk management, and measurable acceptance criteria.
- Drove organization-level adoption: integrated simulation workflows with PLM processes and created a scalable operating model for new programs.
Professional Experience
Head of Simulation Modeling Bureau | Rostselmash
Dec 2021 – Present | Rostov-on-Don, Russia
- Technical Leadership: Scaled a cross-functional simulation engineering team from 2 to 10 members, establishing engineering culture, simulation model review practices, and standardized computational workflows.
- Simulation Workflow Automation: Optimized data preparation and processing across simulation pipelines with Python automation, reducing manual setup time by 12x; integrated workflows with corporate PLM systems (Teamcenter).
- End-to-End Delivery: Led the full lifecycle of complex computational projects (DEM/CFD, MBD/RBD & 1D system simulation) from R&D to serial production validation, ensuring <5% deviation between model predictions and real-world telemetry; simulation-supported H820 harvester awarded AGROSALON Golden Star 2024.
Data Scientist | Chemwatch
Jan 2021 – Jan 2022 | Part-time · Freelance | Remote, Australia
- Built a production batch pipeline for OCR + multilingual NLP to extract structured chemical data (ingredients, CAS numbers, proportions) from 50,000+ MSDS documents across 12+ languages.
- Designed confidence scoring + human-in-the-loop routing with CAS registry validation; reduced manual work by 95% and cut downstream safety analysis time by 40%.
Lead CFD/DEM Engineer | Gomselmash
Aug 2012 – Dec 2021 | Gomel, Belarus
- HPC & Multiphysics Simulation Domain: built CFD/DEM capability from zero (started 2015), trained 3+ engineers, and delivered simulation-validated machinery projects over 9 years, including structural FEA analyses.
- Published calibration methodology (CADFEM Review, 2020) and established verification workflows linking experiments → calibrated parameters → model predictions.
Core Skills
Simulation & CAE
- DEM: Ansys Rocky, Altair EDEM
- CFD: Ansys Fluent, Siemens STAR-CCM+
- FEA: Ansys Mechanical, MSC Nastran/Patran, SolidWorks Simulation
- MBD & 1D: Ansys Motion, MSC Adams, RecurDyn, Simcenter Amesim, MATLAB/Simulink
- Multiphysics: Ansys Workbench
- CAD: SpaceClaim, Siemens NX, SolidWorks, Autodesk Inventor, PTC Creo, AutoCAD, KOMPAS-3D
- Optimization/DOE: Ansys DesignExplorer, Ansys optiSLang
- Post-processing: Ansys CFD-Post, EnSight
- PLM: Teamcenter, Windchill
- HPC & scripting: Python (pipeline automation, Ansys Rocky API), MathCAD
Machine Learning
- Python
- scikit-learn
- XGBoost
- Optuna
- pandas
- NumPy
- Feature Engineering
- Model Evaluation
MLOps
- MLflow
- DVC
- Great Expectations
- Airflow
- Evidently
- Data Quality
- Production ML
Deployment & Production Engineering
- Docker
- Kubernetes
- Helm
- FastAPI
- Celery
- RabbitMQ
- CI/CD
Applied AI & LLM Systems
- LangGraph
- RAG
- LLM
- Qdrant
- Vector Search
- Pydantic AI Tools
Data Engineering
- PostgreSQL
- Redis
- MinIO
- Parquet
Observability & DevOps
- Prometheus
- Grafana
- GitLab CI/CD
- Monitoring
- Observability
- SOPS
- Age
Currently Exploring & Next Steps
Built CAE Copilot (RAG + LangGraph + golden eval) as the Applied AI anchor alongside SoccerPredictAI (MLOps). Currently exploring Stage 3 solver orchestration, async CAE jobs, and LangSmith-based eval extensions.
Solver orchestration Async CAE jobs Simulation surrogates Physics-informed ML LangSmith eval Agent workflows
Education
MSc in Mechanical Engineering — Gomel Technical University P.O. Suhogo, Belarus (2016–2018)
BSc in Mechanical Engineering — Gomel Technical University P.O. Suhogo, Belarus (2007–2012)
Certifications
- LLM: From Understanding to Product — ODS.ai (2026)
- MLOps and Production in Data Science (2.0/3.0) — ODS.ai (2023/2024)
- Machine Learning and Data Analysis — Yandex / MIPT (2020)
- Introduction to Machine Learning — HSE University (2020)
Languages
English: Professional Working Proficiency (B2)
Russian: Native
Interests
3D printing, laser engraving, Arduino, gym.
