ML Engineer
Sigli
Oct 2025 — PresentFine-tuned multilingual CEFR models and built language-processing tools, pipelines, and QA tooling.
• Fine-tuned BERT-based CEFR text classification models for English, Dutch, French, and German, covering all 6 proficiency levels (A1–C2) and improving multilingual scoring consistency.
• Built multilingual training and evaluation pipelines for CEFR prediction, including dataset preparation, feature extraction, error analysis, and comparative model experiments.
• Developed spell-check and autocomplete components using lexicons of up to 200K entries, prefix indexing, and edit-distance candidate generation for user-facing suggestion systems.
AI Engineer
VEON
International telecommunications and digital services group.
Jan 2025 — Oct 2025Delivered a RAG chatbot and document-search pipeline; search fallback reduced the no-answer rate by 37%.
• Built and deployed a stakeholder-facing chatbot on a new RAG stack, delivering 85% answer accuracy with a 5–10 s response time per query.
• Experimented with GraphRAG on Neo4j alongside vanilla RAG, deepening team expertise in graph-augmented retrieval techniques.
• Indexed ~1,000 documents (~40,000 pages) through an OCR pipeline built with pypdfium2 and marker, unlocking semantic search across previously non-digital content.
• Integrated live Bing and Google search fallback, cutting the no-answer rate by 37%, and implemented MCP for end-to-end traceability and maintainable production code.
NLP Engineer
Pivotics
Nov 2023 — Jan 2025Built a car sales timing model and improved Llama 2 inference speed and answer relevance.
• Built a neural network for car sales timing prediction, achieving a 3.8 RMSE versus a 6.5 benchmark.
• Leveraged Nvidia TensorRT-LLM to accelerate Llama 2 inference (7B: 48→106 token/s; 13B: 35→59 token/s).
• Applied prompt engineering to eliminate redundant queries, improving answer relevance by 40%.
Data Scientist
Kaspi.kz
Kazakhstan-based financial technology, payments, and e-commerce company.
Jan 2021 — Oct 2023Built and optimized voice-bot models and monitoring; the bot handled nearly 5 million calls monthly.
• Deployed a high-load BERT-based microservice for an outbound voice bot that classifies client intent, replacing 700 call centre managers and handling nearly 5 million calls per month.
• Ran A/B tests showing a 7 pp error reduction with BERT versus Rasa DIETClassifier.
• Fine-tuned Llama 2 with PEFT to produce personalized, context-aware conversation scripts, and optimized production inference with 8-bit quantization, lru_cache, and JMeter load testing.
• Implemented a CatBoost model for real-time voice bot monitoring, cutting error response times from 24 hours to 1 hour.
