Real-time multi-agent AI monitoring platform for the US trucking industry. Processes thousands of driver–dispatcher Telegram messages daily in EN/UZ/RU via parallel LLM agents. ● Multi-agent architecture: Designed a 5-agent parallel LLM pipeline (Movement, Load & Dispatch, Department, Safety, Behavior) with async fan-out/fan-in — 0.8–2s per-agent latency in production. ● Prompt engineering at scale: Authored 9 production prompts (~95 KB) with strict JSON output, 8-language support, Chain-of-Thought decomposition, few-shot rubrics, enum-locked schemas. ● LLM cost optimization: Implemented Gemini context caching with FIFO eviction and Redis persistence — reduced repeated input costs by ~90%. Built 3-tier cost accounting across 6 Gemini SKUs. ● Provider abstraction: Built pluggable LLM provider layer supporting Google Gemini (cloud) and self-hosted Ollama (Qwen 3.5 35B-A3B MoE) — switchable via single env var. ● Two-phase RAG: Query-planner LLM emits a JSON search plan, answer agent produces cited evidence-grounded responses — significantly reducing hallucinations. ● Fault tolerance: Designed fault-isolated coordinator with custom FailedWithCost exception preserving cost telemetry — single-agent failures never block downstream digital-twin updates. ● High-throughput pipeline: Async FastAPI webhook ingestion handling 100–300 events/sec with HMAC auth, Redis dedup, PostgreSQL idempotent upsert. ● Observability from scratch: Per-call token/cost/latency in PostgreSQL with UUID correlation IDs — end-to-end tracing from webhook to LLM response. ● Production reliability: Jittered exponential backoff on rate limits, primary→fallback model chain, 200-concurrent-call semaphore.

AI solutions for US logistics — FMCSA Hours of Service compliance, driver behavior analysis and real-time voice systems. ● Designed a multi-agent AI system that analyzes driver activity logs and detects FMCSA HOS violations automatically — 11-hour driving limit, 14-hour shift window, 30-minute break, 10-hour rest, 70/80-hour cycle, and PTI checks. ● Built agent-based workflow automation for dispatchers — adding shifts, PTI, overtime checks, violation resolution — accelerated by AI suggestions. ● Designed multi-agent architecture where each agent owns a domain (shift validation, rest-period checks, overtime calculation, cycle-level violations, final report). ● Developed real-time voice translation using OpenAI Realtime API for live two-way driver–dispatcher communication. ● Researched, deployed and benchmarked 5+ speech models (open-source & closed-source) for STT/TTS quality, latency, accuracy and cost. ● Built logistics-domain chatbot with conversational memory; routes complex analytical queries into the violation-detection pipeline. ● Engineered prompt systems with strict JSON output, tool-calling (Calculator, Date & Time) for math accuracy, and robust JSON-parsing fallbacks.

● Led AI team end-to-end — Computer Vision and LLM projects in production. Designed full pipelines: data collection, model selection, deployment, monitoring. ● uzkimyo.uz AI Chatbot (in production) — AI agent that answers user questions over a 200+ table database by auto-generating SQL queries. ● AI HR Assistant Platform — Built an interactive AI system that conducts conversational interview, generates a structured resume, and recommends matching jobs to recruiters (with candidate consent). Parses uploaded resumes via open-source document extraction. ● Safety Monitoring System for Schools/Kindergartens — CV monitoring on surveillance cameras detects physical conflicts (children fighting, staff aggression) and sends real-time alerts. Collected dataset and trained YOLO models independently. ● RAG & Document Search Solutions — Built RAG systems and document-search agents using LangChain and Ollama, integrated with domain-specific data. ● Provided technical leadership: code reviews, model evaluation, technical documentation, team mentoring.

● Worked on freelance Computer Vision and ML projects — object detection, custom dataset training, image processing pipelines. ● Took time for travel, self-development, and continuous learning in the latest AI/LLM technologies (RAG, AI agents, multimodal systems, prompt engineering). ● Studied production-grade AI architectures and the modern LLM ecosystem — staying ahead of rapid industry changes.

● Taught AI and Machine Learning to two student groups (~20 students total) with an individualized approach. ● Designed and delivered a full curriculum: Python fundamentals, data visualization (Matplotlib, Seaborn), NumPy, Pandas, OpenCV (Computer Vision), and PyTorch (ML). ● Prepared all course materials from scratch — lecture slides, code examples, and practical homework for each module. ● Course projects (with students & team): Vehicle Detection & License Plate Recognition (YOLO + OCR pipeline), Road Lane Detection for driver assistance scenarios, Custom dataset of 10,000+ images annotated and trained on YOLO.

Real-time video analytics for industrial clients (mining, manufacturing). ● PPE Detection — Model that detects whether workers at mining enterprises wear helmets and special uniforms via CCTV video streams. ● Distance Measurement — CV solution to measure how far a person stands from the camera in real time. ● Product Counting — CV system that automatically counts products moving through containers on production conveyors. ● Worked on the full ML pipeline: data collection, annotation, model training, and deployment to production.