Azizbek Siddikov

• ▸Accomplished 15–20× faster YOLOv11 training by profiling input pipeline bottlenecks, building an offline 4K→640 preprocessing flow, and migrating NAS datasets to local GPU storage; cut epoch time 3–3.5 h → 11–14 min for 600K images; mAP@0.5:0.95 97%.
• ▸Shipped a company's knowledge assistant with citations by integrating Gemini File Search, implementing SSE streaming, and building an admin console for ingestion lifecycle, error auditing, and conversation traceability.
• ▸Refactored the order page into a modular config interpreter for pricing, shipping, promos, and tax, cutting duplicated code ~4× and enabling faster and safer launches via centralized validation.
• ▸Automated hardware QA throughput by converting calibration/PCB validation into a batch-mode executable with structured logs and failure codes; reduced operator steps 8→2 and supported 8–10 devices per run in parallel.
• ▸Developed an LLM-based Text-to-SQL agent-style AI workflow engine so the agent dynamically generates and safely executes SQL queries, enabling self-service analytics for non-engineering teams.
• ▸Operated and managed AWS-based cloud infrastructure, supporting backend services, storage, and compute resources for the platform.

• ▸Built and launched a "Twitter for books" social platform end-to-end with a 3-person team, onboarding 70+ early users and delivering core loops (auth, posting, follows, feed, search).
• ▸Engineered the feed + GraphQL API for performance and scale using cursor-based pagination and resolver batching/caching to eliminate N+1 patterns and reduce backend load.
• ▸Containerized the project and introduced a CI/CD pipeline including validation and logging for faster, more stable deployments.

• ▸Built and optimized dataset ingestion/QA pipelines for CV image/video projects, automating validation (label–file consistency, format checks, outlier stats) and surfacing quality metrics via dashboards.
• ▸Developed an internal Python library (preprocessing + file/manifest management + converters), reducing dataset preparation time 60% and standardizing outputs across projects.
• ▸Led 3 data labelers by defining annotation standards and review workflows across CVAT and MakeSenseAI, improving labeling consistency and reducing rework.
• ▸Developed a LangChain-based LLM data cleansing agent that analyzes table data quality and automates cleansing (missing value handling, duplicate detection, data type correction, outlier removal), providing change summaries and data quality statistics.

Full-stack events platform that web crawls events using AI, batches processings, and GraphQL API.

Full-stack online fashion store where users can browse products, place orders, and manage purchases.

Upload documents, ask questions, get answers grounded in your files.

AI translation of medical documents considering ICD codes.

World ID–verified Q&A platform: post questions, share answers, accept best response; bot and Sybil protection.

Connected device with on/off controls; web interface for interaction and monitoring.

Detect and track vehicles with SORT, detect and read plates with custom YOLO + EasyOCR; CSV export and visualization.

Preprocess text from CSV; run OCR with Tesseract and EasyOCR; compare similarity and processing time.

EDA and prediction of resale flat prices (2017–2024): preprocessing, feature engineering, regression models (Random Forest, XGBoost, etc.).

Solve a murder mystery using SQL from a problem sets of Harvard CS50 lectures.