Ahsan Mustafa

• 01 · Platform50% faster reports

  Enterprise data lakehouse

  More than ten sources flowing through Bronze, Silver and Gold layers on Databricks and Delta Lake, with self-serve BI on top.

  DatabricksDelta LakeAzure Data FactorySnowflakeAirflow

• 02 · Streaming60% lower latency

  Real-time financial streaming

  CDC and event streaming for millions of transactions a day, feeding near-real-time analytics and anomaly detection.

  KafkaSpark StreamingRedshiftPythonCDC

• 03 · ResearchmAP50 55% to 86%

  Road defects from the sky

  YOLOv12s and YOLO26 fine-tuned on drone imagery to find six kinds of pavement damage. Published and presented in 2026.

  YOLOPythonRoboflowPandas

• 04 · Open projectIn progress

  Carbon vs. Capital

  Emissions data linked to US-listed companies, built end to end with Azure Data Factory, Databricks, Snowflake and Power BI.

  Azure Data FactoryDatabricksSnowflakePower BI

• 05 · AI agents4 workflows, 1 bot

  A personal AI agent on n8n

  A Telegram agent that advises on a career track, digests news and job postings, and builds study plans from live web search. n8n and GigaChat on Cloud.ru, built for School 21.

  n8nGigaChatTelegram Bot APITavilyHH.ru APIDocker

  Read the write-up →

1. [1]

   Detection of Road Defects from Unmanned Aerial Vehicles Using Machine Learning

   A. Mustafa, M. Yu. Kataev, D. S. Efremenko · Light & Engineering, vol. 34, no. 3, pp. 4 to 12, 2026

   Google Scholar ↗Copy BibTeX

   Abstract+

   The timely detection of road surface deterioration is a critical factor in infrastructure maintenance and traffic safety. Conventional inspection methods remain labour-intensive, costly, and prone to human error. In contrast, unmanned aerial vehicles (UAVs) combined with advanced computer vision techniques provide a scalable and cost-effective alternative. In this study, we propose a neural network model based on the YOLOv12 object detection framework for identifying road surface defects. The model is trained on a UAV-PDD2023 dataset containing over 2400 high-resolution aerial images annotated across six categories of road defects. Two model configurations with varying preprocessing steps and hyper-parameters are evaluated to assess robustness and generalization. The evaluation of the model efficiency was performed when RGB imagery is replaced with single-channel counterparts. Results show that the accuracy of defect detection remains largely unaffected: over 90 percent of the defects detected in RGB imagery were also correctly identified in single-channel images, with a performance variation of less than 5%. This indicates that a single trained model can be effectively applied to images with different spectral characteristics.

2. [2]

   Structured Training Ablation and GSD-Adaptive Orthophoto Inference for UAV-Based Pavement Defect Detection with YOLO26

   A. Mustafa, M. Yu. Kataev, D. S. Efremenko · Moscow University Physics Bulletin, 2026, in press · Presented at DLCP 2026

   DLCP 2026 ↗Copy BibTeX

   Abstract+

   Automated detection of pavement surface defects from unmanned aerial vehicle (UAV) imagery can substantially reduce the cost and labour of road condition surveys. Two questions that remain largely unanswered in current literature are which training configuration choices actually drive accuracy gains, and how well a model trained on low-altitude UAV images generalises to the higher-altitude orthophotos used in real-world deployment. This paper addresses both. We train YOLO26 on the UAV-PDD2023 benchmark dataset (2,439 images, six defect classes) and run a controlled five-configuration ablation study that independently varies hyperparameter settings, preprocessing, and data augmentation. The best combination reaches mAP50 = 0.859 and mAP50:95 = 0.631 on the held-out test split, a 57% and 133% improvement over the default baseline respectively. The hyperparameter configuration, which bundles increased input resolution with extended training to 300 epochs, is the single largest contributor; preprocessing and augmentation each add complementary gains. We note that this configuration combines resolution and epoch changes, so the resolution effect is not fully isolated in the current design. We then deploy the best model on three real-world GeoTIFF orthophotos through a ground-sampling-distance (GSD) adaptive tiling pipeline. The pipeline selects tile dimensions to match the ground coverage seen during training, applies cross-tile non-maximum suppression, and converts bounding boxes to WGS84 coordinates. Despite a measurable domain gap, the pipeline identifies all six defect classes across scenes spanning up to 322 × 843 m, with 2,201 detections on the largest mosaic. Exported GeoJSON and annotated GeoTIFF outputs are validated in QGIS against source rasters.

All my papers on Google Scholar ↗

• 2024 to 2026

  M.Sc. Computer Engineering (Data Science)

  National University of Science and Technology MISIS

  Graduated with honours (Red Diploma)CGPA 5.0 / 5.0

• 2025 to 2027

  Data Science Program

  School 21 by Sber

  Winner, Intensive Pool 2025100%

• 2015 to 2019

  B.Sc. Computer Science

  National University of Computer and Emerging Sciences (FAST)

  Silver MedalistCGPA 3.6 / 4.0

Certifications

• VK Education
  2025

  Practical project: predicting social-network users' gender

  ID 480-418972

• Google
  Sep 2022

  Google Analytics Certification
• C

  Cloud.ru and School 21
  Dec 2025

  Cloud.ru Evolution Fundamentals

  ID 4817

Awards

• 2026Diploma of High Achievement, Computer and Data ScienceOpen Doors olympiad, Global Universities Association
• 2026Master's degree with honours (Red Diploma)NUST MISIS
• 2025Prize-winner, Applied Mathematics and AIOpen Doors olympiad, Global Universities Association
• 2025Winner, School 21 Intensive PoolSchool 21 by Sber
• 2023Winner, Open Doors olympiad: Russian Government Scholarship, Computer and Data Science trackGovernment of the Russian Federation
• 2019Runner-up, ACM ICPC 2019 regional contestACM GIKI Chapter, Pakistan
• 2019Silver Medalist, BS(CS) semesters VI and VIIFAST