Development within a Data Science and machine-learning team. The portfolio covers industrial combinatorial optimization, natural-language processing, reinforcement learning, recommender and multi-agent large-language-model systems — from applied prototype to production. Mentoring of a student intern.
Product-shipment optimization in in-plant logistics
Combinatorial scheduling of loading under limited resources. An own solver built on a “slotted knapsack” abstraction (branch and bound, heuristics), integrated with 1C; covers the bulk of planning factors.
Outbound delivery-logistics optimization
Assignment of carriers and loading time-slots to orders: a mixed-integer programming model (Pyomo, HiGHS) and a genetic algorithm (DEAP) with lexicographic multi-objective optimization.
Corporate-correspondence classification
Multi-class routing of letters to recipient units under strong class imbalance and personal data. Fine-tuning of BERT-family models, text extraction from scans via a vision-language model, augmentation, hyperparameter search.
LLM assistant for business processes
Process digitalization via large language models with structured output (LangChain, LangGraph).
RL agent for reservoir-pressure maintenance
Adapting the experience buffer for parallel data collection by several workers on Windows, for a reinforcement-learning agent (PPO) controlling waterflooding in a hydrodynamic simulator.
Corporate-learning recommender system
Design of a hybrid recommender over a competency model: feature fusion, cold start, loss function, quality metrics.
Digital-twin orchestrator (multi-agent system)
Conceptualizing business cases and mapping them onto the orchestrator’s mechanics, working out the multi-agent-system architecture: a system model as an RDF graph (SPARQL, scenario versions), an ontology in OWL and SHACL, agent coordination via ask-tell and rollout mechanics.
Pre-project backlog generation (LLM, RAG)
Working out an approach to generating a task backlog for the pre-project stage via a large-language-model and retrieval-augmented-generation pipeline: applicability analysis and solution-architecture design.
Team-planning dashboard
An internal dashboard of load, full-time-equivalent bookings and absences on OpenProject: Docker Compose, PostgreSQL, backup and restore, integrations.
Review of satellite-imagery computer-vision projects
Technical review: reconciling reports with code, checking reproducibility and metrics, code-duplication analysis, review of inference infrastructure (FastAPI, S3, JWT).
Mentoring
Mentoring of a student intern: review of the optimization model and data parsing, work planning.