Applied AI Engineer focused on reliable LLM systems, RAG, document intelligence, and Python services.
I build AI applications across the full pipeline: data and document ingestion, retrieval, structured LLM outputs, validation, evaluation, backend integration, and deployment. I am especially interested in systems where model results must be traceable, testable, and useful beyond a demo.
- LLM agents for complex documents and weakly structured data
- RAG, semantic search, reranking, and retrieval evaluation
- Structured outputs, quality gates, caching, and failure handling
- Applied ML and NLP experiments
- Python backends and end-to-end AI products
| Project | What it demonstrates |
|---|---|
| Viva AI | Adaptive oral exam coach grounded in the learner's materials. Uses Pydantic-validated LLM outputs, deterministic criterion-based scoring, local Ollama inference, and grading regression tests. Video demo. |
| ProofStudio | Traceable generative media pipeline with provenance manifests, content-addressed storage, SHA-256 verification, retries, CI, Backblaze B2, and Cloudflare Workers AI. Live application. |
| Video RAG | Multilingual video-fragment retrieval with scene detection, ASR, VLM captions, dense/sparse/ColBERT retrieval, FAISS, fusion, and reranking. |
| City Route Planner | Production web service for walking routes across 90 Russian cities. FastAPI, React, Leaflet, OpenStreetMap data, OSRM/Valhalla routing, caching, Docker, and operational health checks. |
| MOEX Analytics | End-to-end market-data ETL and analytics with Prefect, Dask, Pandas, Streamlit, and Plotly. |
In my professional work, I develop LLM-based systems for analyzing complex documents and business data. My recurring engineering concerns include robust document loading, OCR fallbacks, structured extraction, evidence-based validation, conservative review states, deduplication, retries, and reproducible reports.
Public descriptions of this work are intentionally generalized and contain no internal data, prompts, infrastructure details, or client documents.
I also explore cases where adding more AI does not automatically improve the result. In a 2026 coursework experiment on CommonsenseQA and ConceptNet, a GNN-based retriever substantially improved retrieval Hits@3 over a vector-search baseline, while the final LLM answer accuracy decreased. This became a useful case study in evaluating the complete RAG pipeline instead of retrieval alone.
Python · FastAPI · Pydantic · LangChain · LangGraph · PyTorch ·
FAISS · Pandas · SQL · Docker · GitHub Actions
I am developing deeper expertise in reliable AI agents, retrieval evaluation, document intelligence, and production-oriented ML systems. I am interested in Applied AI, LLM/NLP, ML Engineering, and AI Backend opportunities.