"𝔅𝔲𝔦𝔩𝔡𝔦𝔫𝔤 𝔱𝔥𝔦𝔫𝔤𝔰 𝔣𝔯𝔬𝔪 𝔰𝔠𝔯𝔞𝔱𝔠𝔥 𝔲𝔫𝔱𝔦𝔩 𝔱𝔥𝔢𝔶 𝔴𝔬𝔯𝔨. 𝔚𝔢 𝔞𝔯𝔢 𝔫𝔬𝔱 𝔶𝔬𝔲𝔯 𝔨𝔦𝔫𝔡𝔰."
ML Engineer focused on LLM systems, AI agents and deep learning.
MIREA — Artificial Intelligence & Machine Learning. Hackathon finalist. Building transformers on JAX for fun.
I build end-to-end ML systems — from raw data and feature engineering to production inference and deployment. Currently deep into LLM internals, agent architectures and RL theory.
Production RAG system (~7000 queries) optimized for BERTScore-Recall-L
- FAISS (bge-m3) + BM25 + RRF · cross-encoder reranking
- vLLM continuous batching · 2×T4 GPU-aware architecture
- Latency reduced by ~25%
- → github
ML pipeline for personal data leak detection
- 3-tier architecture: rule-engine → classification → scoring
- TF-IDF + LogReg · CatBoost with contextual scoring
- 35% FP reduction
- → github
Production CV system for price tag extraction from video stream
- YOLOv8 + custom IoU tracker · PaddleOCR → EasyOCR cascade
- Throughput: 7 → 19 FPS · OCR inferences reduced by 22%
- → github
Predictive model for store RTO forecasting · 90.45 score
- Time-series feature engineering · gradient boosting · overfitting control
- → github
Byte-level BPE tokenizer · GPT-2 regex · O(n log n) complexity
- Built for deep understanding of LLM tokenization internals
- → github
Transformer implementation: Self-Attention · Multi-Head · MoE · MLA
- Researching memory efficiency and inference optimization
- RAG experiments on custom architecture — ongoing
stack = {
"ml / dl": ["PyTorch", "JAX", "CatBoost", "XGBoost", "Scikit-learn"],
"llm / nlp": ["vLLM", "LangChain", "LangGraph", "FAISS", "BM25"],
"infra": ["Docker", "FastAPI", "SQL", "Git"],
}