Applied AI & MLOps

I focus on the engineering side of AI: building reproducible, maintainable systems that work in production — not just in notebooks. My work spans the full lifecycle: data ingestion, quality control, feature engineering, experiment tracking, model registry, serving, monitoring, and CI/CD — from classical ML pipelines to LLM/RAG systems with agent routing, mandatory citations, and golden evals.

Projects

CAE Copilot – Engineering AI Copilot

May 2026 – Present | cae-copilot.dmitryivanov.dev

LangGraph RAG FastAPI Qdrant Streamlit

Live Demo Docs API (Swagger)

Engineering AI Copilot for technical documentation, calculations, and traceable answers — an Applied AI portfolio project built with the same production mindset as SoccerPredictAI (tests, docs, CI, k8s).

Not solver automation. Live MVP: document-grounded RAG, engineering calculation tools, LangGraph routing, and structured reports with mandatory citations.

Public key numbers

Metric	Value
CAE docs corpus	9 manuals ingested
Maturity stage	Stage 2 shipped (RAG + citations, LangGraph, tools, eval)
Golden Q&A eval	14 cases; offline baseline 100% intent / citation / tool
Deployment	Separate FastAPI + Streamlit images, Qdrant, Helm/k8s

Architecture

Engineer → Streamlit UI → FastAPI → LangGraph
                              ├─ RAG (Qdrant + citations)
                              └─ Engineering tools (beam, FoS, bolt joint, sections)

Traceable demo path: Question → Retrieval → Citation → Calculation tool → Report

Evidence & docs

Problem statement & positioning: docs.cae-copilot.dmitryivanov.dev/portfolio/problem-statement
Architecture overview, data flow, and demo walkthrough: portfolio overview
Live demo: cae-copilot.dmitryivanov.dev
API (Swagger): api.cae-copilot.dmitryivanov.dev/docs

Tech Stack

Layer	Tools
Agent orchestration	LangGraph, Pydantic tool schemas
RAG	Qdrant, multilingual-e5-base, page-aware chunks
Serving	FastAPI, Streamlit
Infrastructure	Docker, Kubernetes, Helm
Quality	Golden Q&A eval, lint, typecheck, pytest
Docs	MkDocs

Key Engineering Decisions

Traceable knowledge workflows: retrieval → citation → calculation tool → report
Reliability-first: deterministic Python tools, golden Q&A eval, citations, offline fallback
Production layout: separate FastAPI and Streamlit images, Qdrant vector store, k8s Helm chart

Evidence pack (fast links)

For interviews, I use these pages to demonstrate traceable agent design: RAG with citations, tool routing, eval discipline, and production deployment.

Portfolio overview Problem statement Architecture overview Data flow Runtime view Demo walkthrough Status (Stage 2 shipped) Deployment API (Swagger)

SoccerPredictAI – End-to-End MLOps Platform

May 2025 – May 2026 | soccer.dmitryivanov.dev | GitLab

MLOps Kubernetes XGBoost FastAPI MLflow DVC

Live Demo Docs (Evidence) GitLab

A production-style pet project built to demonstrate the full ML lifecycle in a real domain (football match prediction). Deliberately engineered as a system, not a collection of notebooks — domain-agnostic proof of production ML/MLOps discipline.

Public key numbers

Metric	Value
Holdout log-loss	1.006 (bookmaker benchmark ~0.97)
Holdout ROC AUC	0.643
Calibration (ECE)	0.004
Holdout set size	135,970 matches (2024+)
DVC pipeline stages	22
Test suite	560+ passing tests
Infrastructure cost	<€30/month (single self-hosted VPS)

Evidence & docs

System overview, readiness matrix, results, and demo script live in the project docs: docs.soccer.dmitryivanov.dev
End-to-end pipeline analysis reports (Quarto): Pipeline summary & stages 01–07
Reproducibility golden path (dvc pull → dvc repro → mlflow ui): see Quickstart in the docs

Pipeline & Architecture

Scraping → PostgreSQL/MinIO → Great Expectations (DQ)
    → Feature Engineering (rolling stats + ELO ratings)
    → Temporal CV screening (5 models × 6 data fractions, MLflow)
    → Optuna hyperparameter tuning (CV-only, no holdout leakage)
    → Final retrain on full dataset → Model Registry
    → FastAPI (sync) + Celery/RabbitMQ (async inference)
    → Kubernetes + Helm → Prometheus + Grafana + Evidently (drift, in progress)

Tech Stack

Layer	Tools
Orchestration	Airflow, DVC
Data & Storage	PostgreSQL, MinIO (S3), Parquet
Data Quality	Great Expectations
Feature Engineering	pandas, NumPy, scikit-learn
ML / Tuning	XGBoost, scikit-learn pipelines, Optuna
Experiment Tracking	MLflow (tracking + model registry)
Config	Hydra, params.yaml
Serving	FastAPI, Celery, RabbitMQ
Infrastructure	Docker, Kubernetes, Helm
Monitoring	Prometheus, Grafana, Evidently
CI/CD	GitLab CI, pre-commit, SOPS + Age
Docs	MkDocs, Quarto

Key Engineering Decisions

Anti-leakage experiment design: model selection by CV logloss only; holdout evaluated once in a dedicated `final_train` stage
MLflow nested runs: parent run per data fraction, child run per model — full comparison matrix in the UI
DVC pipeline: all stages versioned with input/output hashes; reproducing any experiment is a single command
Data quality gates: Great Expectations suites at raw, interim and features layers — pipeline fails on schema violations
Temporal CV: walk-forward folds with `startTimeUtc` boundaries — no future leakage in cross-validation
Kubernetes deployment: Helm chart, SOPS + Age secrets, GitLab CI quality gates on every push

Evidence pack (fast links)

For interviews, I use these pages to demonstrate production reality: reproducibility, architecture, quality gates, monitoring, and results.

Docs home Quickstart (dvc pull → repro) Demo walkthrough Results Pipeline reports (Quarto) System overview Implementation status Monitoring CI/CD quality gates

Chemwatch – ML Pipeline for Multilingual Document Processing

Jan 2021 – Jan 2022 | Freelance | Remote, Australia

NLP OCR Python

Problem: Chemwatch (AU) maintains a large database of chemical Safety Data Sheets (SDS/MSDS) in 20+ languages. Manual extraction of structured data (ingredient names, CAS numbers, proportions) was the bottleneck.

Approach:

OCR pipeline to handle both text-native and scanned PDFs
Multilingual NLP (rule-based + ML classifier) to locate the ingredients section across document formats
Confidence scoring + human-in-the-loop filtering for uncertain cases
Validation against known CAS registry

Result: Reduced manual data extraction workload by 95%. Deployed as a batch processing pipeline.

Key Engineering Decisions

Processed 50,000+ documents in EN, DE, FR, ES, IT, NL, SV, PT, FI, SK, VI and more
Built custom section-detection model handling non-standard SDS formats and OCR noise across 20+ document layouts
Reduced manual extraction workload from ~8h/day to <30min/day (95% reduction) via confidence-scored human-in-the-loop filtering and CAS registry validation
Stakeholder dashboards cut chemical safety analysis time by 40%; delivered as a scheduled batch pipeline

Sample annotated MSDS documents (section detection output)

Skills

Core ML: scikit-learn, XGBoost, Optuna, feature engineering, temporal CV, model calibration MLOps: MLflow, DVC, Airflow, Great Expectations, Evidently, Prometheus, Grafana Applied AI: LangGraph, RAG, Qdrant, LLM integration, golden eval, citation-grounded reports Serving: FastAPI, Celery, RabbitMQ, Docker, Kubernetes, Helm Data: PostgreSQL, MinIO, pandas, NumPy, Parquet Engineering: Hydra configs, GitLab CI/CD, SOPS + Age, pre-commit, MkDocs

Certifications & Courses

LLM: From Understanding to Product — ODS.ai (2026)
MLOps and Production in Data Science 2.0 — ODS.ai (2023)
MLOps and Production in Data Science 3.0 — ODS.ai (2024)
Machine Learning and Data Analysis Specialization — Yandex / MIPT (2020)
Introduction to Machine Learning — HSE University (2020)

Coursera certificates for all 6 courses of the ML and Data Analysis specialization (Yandex / MIPT), Introduction to Machine Learning (HSE University), and Introduction to Data Science.