avazbek.xyz
Open to remote contract work

Avazbek Isroilov

Big Data / ML Engineer

I build production ML systems and large-scale data pipelines — specializing in distributed graph processing, computer vision architectures, NLP & vector search clustering, and high-throughput geospatial & GIS data processing workflows.

Get in touch View case studies

Selected Work

Galaxy of Science

Distributed Citation Graph Processing & 3D Visualization
Live Demo ↗
DISTRIBUTED GRAPH METRICS science-galaxy.avazbek.xyz
112M
Scientific Papers
2.5B
Citation Edges
60 FPS
WebGL / GLSL Render
Problem
Analyzing global scientific influence across massive citation topologies requires computing iterative graph algorithms over hundreds of millions of nodes and billions of edges, which exceeds single-node memory and compute capabilities.
Approach
Engineered a distributed PageRank processing pipeline operating over 112 million research papers and 2.5 billion citation edges. Standardized node embeddings and projected the high-dimensional topology into a continuous 3D coordinate space rendered via custom WebGL/GLSL shaders.
Result
Enabled real-time, interactive exploration of global academic literature clusters at 60 FPS directly in the browser, providing instant visual citation traversal across massive datasets.
Distributed Graph Processing PageRank Big Data Pipeline WebGL / GLSL Three.js

Vegetation Segmentation at Scale

GeoAlert Automated Data Annotation & CV Model Scaling
Official Release ↗
Problem
Individual tree crown segmentation models suffered from severe accuracy bottlenecks across diverse biomes due to manual labeling constraints and scarce ground-truth data.
Approach
Devised an automated weak-supervision and data-annotation pipeline to synthetically bootstrap and clean training datasets at scale for high-resolution satellite imagery across four continents.
Result
Boosted GeoAlert's global Forest & Trees vegetation model performance significantly across multi-continental validation sets: F1 score improved from 0.513 to 0.850 and Intersection over Union (IoU) increased from 0.396 to 0.744.
PyTorch Computer Vision Semantic Segmentation Automated Annotation Geospatial CV

Urban Heat Island Downscaling

Cross-Modal Attention U-Net & Production MLOps Stack
Problem
Coarse-resolution thermal satellite channels lack the spatial granularity required for urban microclimate policy and targeted Heat Island mitigation in rapidly developing metropolitan regions like Tashkent.
Approach
Built a dual-encoder U-Net architecture featuring cross-modal attention mechanisms to fuse heterogeneous multi-source satellite imagery. Downscaled land surface temperature to 30m spatial resolution, rigorously validated against Wald's protocol and traditional TsHARP / DisTrad baselines.
Result
Deployed an end-to-end automated MLOps pipeline featuring model tracking, orchestration, data validation, and continuous deployment workflows.
PyTorch Cross-Modal Attention MLflow Prefect CI/CD Model Monitoring

MethaneWatch

Automated Methane Plume Detection via Attention U-Net
Live App ↗
Problem
Point-source greenhouse gas emissions produce faint spectral absorption signatures in satellite imagery, making automated real-time leak detection challenging across vast spatial swaths.
Approach
Implemented a deep learning pipeline leveraging Attention U-Net neural networks to isolate spectral bands and segment high-concentration methane plumes from background noise.
Result
Established an automated detection baseline capable of flagging emission anomalies across multi-spectral satellite passes with minimal false positive rates.
Attention U-Net PyTorch Multi-Spectral CV Data Pipelines

Technical Stack

Machine Learning & CV
  • Python & PyTorch
  • Semantic Segmentation & U-Net
  • BERT, ByT5 & Gemma Fine-Tuning
  • SentenceTransformers & Vector Spaces
  • Weak Supervision & Auto-labelling
Data Engineering
  • Apache Spark & PySpark
  • SQL (PostgreSQL)
  • Distributed Graph Processing (PageRank)
  • minHash + LSH Deduplication
  • Scikit-Learn & Cluster Analysis
  • ETL / ELT Web Scraping Pipelines
Geospatial Specialization
  • Google Earth Engine
  • Landsat / Sentinel / MODIS
  • ArcGIS API & Mapbox
  • Diffusion-based Super Resolution
  • Multi-Modal Satellite Fusion
MLOps & Infrastructure
  • LLM Agent Deployment
  • MLflow & Prefect Orchestration
  • Apache Airflow
  • GitHub Actions & CI/CD
  • Model Drift Monitoring
  • Docker & Microservices Architecture

Professional Experience

Tashkent | Dec 2025 – Present
  • Architected and deployed an end-to-end data pipeline utilizing OCR models for automated text extraction from millions of scanned enterprise documents.
  • Trained and fine-tuned small-scale NLP models (BERT, ByT5, Gemma) via instruction tuning to clean OCR artifacts and reconstruct structural text integrity.
  • Engineered document-level and line-level deduplication pipelines leveraging the minHash + LSH algorithm, identifying and eliminating ~2M duplicate documents out of ~5M total records.
  • Developed an automated metadata standardization pipeline in Python (SentenceTransformers, Scikit-learn) processing 16.2M text occurrences, achieving a 99.86% reduction in topic cardinality (712K → 1K) via vector space clustering, memory-optimized parallel processing, and dynamic domain majority voting.
RnD Engineer in Computer Vision
GeoAlert ↗
Tashkent | Aug 2025 – May 2026
  • Worked in a fast-paced startup specializing in GIS and AI, driving research and production scaling of cutting-edge Computer Vision architectures for satellite and UAV imagery.
  • Contributed core R&D to GeoAlert’s flagship global Forest & Trees vegetation model, whose 2026 release achieved an F1 jump from 0.513 → 0.850 and IoU from 0.396 → 0.744 across 4 continents (view official release announcement ↗).
Tashkent | Feb 2025 – Nov 2025
  • Developed Diffusion-based Super Resolution models for satellite imagery reaching a PSNR of 26–28dB.
  • Conducted satellite image analysis via ArcGIS API and constructed robust automated training/inference pipelines for solid waste and tree crown detection segmentation models.
  • Engineered and deployed autonomous LLM agents as scalable, containerized microservices.

Education & Academic Background

BSc Computer Science
University of Warwick
Coventry, UK | Oct 2021 – Jun 2024
  • Degree Classification: Second Class
  • High-scoring modules: Computational Physics (80%), Software Engineering Project (77%), Mathematics for Computer Scientists (75%), Logic & Verification (72%), Database Systems (65%).
  • Final Year Project: Personal Calisthenics Mobile Application Assistant leveraging Google's BlazePose pose estimation library.
GCE A-Levels
Bellerbys College
Brighton, UK | Jan 2020 – Jun 2021
  • Grades: A*A*A*A* (Mathematics, Further Mathematics, Physics, Computer Science).
  • Nominated as the Best Computer Science Student.
  • Final Year Project: Text recognition software built using TensorFlow Keras.

Let's Work Together

Available for remote contract work, consulting on ML infrastructure, GIS/remote sensing data pipelines, large-scale NLP/OCR pipelines, vector clustering, and distributed graph processing.

avazbekisroilov2002@outlook.com