Aleksei I.

Machine Learning and AI Engineer. 8+ years shipping applied AI into production: LLM agents and RAG, NLP, speech, computer vision, and CUDA. I build systems that hold up in production, not demos, and every claim below is a number I measured. LLM and NLP - Semantic CV-to-job matching for an EU job board: 10,001 CVs against 10,003 vacancies, 100M cosine pairs in about 18s on one RTX 4090 (bge-m3), then a two-stage LLM re-rank that spreads the top 10 from 0.93 down to 0.71 where the raw cosine moves only 0.03. - LLM CV builder: messy PDF / DOCX / LinkedIn text into a strict Pydantic record bound to the client's 674-value taxonomy, so invalid output is structurally impossible. Emits a recruiter-ready LaTeX PDF and database-ready JSON. - Crypto signals from tweet embeddings (live client program): Qwen3-Embedding-8B on an A40, 1M+ tweet vectors in Qdrant, 44 symbols scored every 60 seconds, 4.65M events ingested, 10-day zero-error run. Speech - On-prem call-centre speech-to-text, in production 2020-2022, feeding fraud detection and call routing. Benchmarked on hand-labeled calls: Whisper 20% median WER, Vosk 28%, Google 32%. - Diarized transcription toolkit, TTS, and cloned-voice agents. Computer vision and CUDA - Multi-camera 3D capture for film relighting: 33 datasets delivered, 3-32 synchronized cameras each, GPU space carving at 1.98 s/frame versus 65 s in vectorized NumPy (about 33x). 5.0 Upwork review. - GPU brute-force passphrase recovery: about 58,000 candidates/s on one RTX 4090, 160x over CPU, sharded across multiple GPUs and machines. - Open-set retail robot vision on ROS 2; banknote detection at the cash desk. YOLO, Grounding DINO, RealSense. How I work Anti-hallucination is enforced in the code path, not in a prompt: strict schema validation, retry on violation, closed enum vocabularies. Success criteria are pre-registered before results exist, models are judged on time-ordered splits, and negative results get reported rather than buried. Stack - LLM and agents: OpenAI API, Claude Agent SDK + MCP, LangChain / LangGraph, RAG, vLLM, Hugging Face Transformers, fine-tuning, structured outputs with Pydantic + instructor, hand-written tool-use loops, chatbots on Telegram and web - NLP and search: embeddings, semantic search, vector databases (Qdrant), clustering, topic discovery, classification, multilingual pipelines - Speech: Whisper, Vosk / Kaldi, Deepgram, Google Cloud STT, TTS, speaker diarization - Vision: OpenCV, YOLO, Grounding DINO, TensorFlow Object Detection, RealSense RGB-D, ROS 2, NVIDIA Jetson - Classic ML: PyTorch, TensorFlow, scikit-learn, CatBoost, forecasting, time-series validation, reinforcement learning - GPU: CUDA C++, PyCUDA, Nsight Compute, multi-GPU sharding - Data and infra: Python, FastAPI, Docker Compose, PostgreSQL, Redis, BigQuery, pandas / NumPy / Parquet, n8n, Linux VPS, GCP Full case studies for each project above, with architecture diagrams and honest results including what did not work, are in my Portfolio section below. Tell me what you are building and I will tell you plainly whether I am the right fit for it.