01 / Intro

Mohamed Hendawy

Senior Computer Vision Engineer, Machine Learning Engineer, and Software Engineer.

I build real-time perception systems for robotics and video analytics. The work covers object detection, segmentation, multi-object tracking, model evaluation, data pipelines, and production debugging, in Python and C++.

I have more than four years of this work. I pay attention to hard cases, dataset quality, and whether the software around a model can be run again by someone else.

Now

  • AIVIS Studio, Senior Computer Vision Engineer, part-time, remote
  • Swap Robotics, Computer Vision Engineer and Software Engineer, full-time, remote
Portrait of Mohamed Hendawy

02 / Experience

Selected experience

Two current roles, then earlier work. The current roles overlap. AIVIS Studio is part-time. Swap Robotics is full-time.

April 2026 to present

AIVIS Studio

Senior Computer Vision Engineer

Part-time, remote

I helped plan and build, with the team, a real-time computer vision system for kitchen safety and hygiene monitoring. Hands-on work includes object detection, person tracking, temporal stabilization, and service contracts for downstream violation processing.

I train and evaluate models, keep annotation workflows and dataset quality in order, and write scenario-based regression tests, including hard-negative cases. I keep experiments reproducible and put CI checks around the work.

I supervise two computer vision engineers, review their work, and plan tasks with the team lead. I remain hands-on.

June 2023 to present

Swap Robotics

Computer Vision Engineer | Software Engineer

Full-time, remote

I build real-time perception and robotics software for autonomous outdoor vehicles, with the team. That includes object detection, segmentation, multi-object tracking, and multi-camera obstruction monitoring.

I work on camera orientation using roll, pitch, and yaw signals, autonomy and teleoperation safety behavior, remote-standby fallbacks, motion and attachment controls, and operator diagnostics. I also work on LiDAR mapping and path planning, and on ROS and Gazebo simulation for articulated robots.

I built annotation and segmentation workflows using SAM and YOLO.

November 2022 to July 2023

Devservers

Machine Learning Engineer and Computer Vision Engineer

I worked on anomaly detection with LSTM and attention mechanisms, and on a computer vision near-miss detection system.

October 2021 to March 2022

Autotronics Research Lab

Computer Vision Algorithm Engineer Intern

I worked on autonomous-vehicle perception, CNN architectures, and implementing ideas from scientific papers.

03 / Projects

Selected projects

Two public projects in more detail, then two compact records. Production systems stay with their teams.

01

AugmentLens

Python, OpenCV, Albumentations

View repository

December 2025

Problem

Image augmentations can damage training data without raising an error. Bounding boxes can leave the frame, masks can stop matching the image, and transforms can change a sample in ways that stay hidden until training quality drops.

Build

AugmentLens wraps an existing pipeline and captures the sample before and after a run. That makes corrupted boxes, masks, or transformations visible before they reach training.

Evidence

Public documentation describes Albumentations support, automatic bounding-box clamping, a pass-through production mode, multiple tensor and array layouts, and mask-integrity inspection.

Inspection path
  1. Input frame
  2. Augmentation pipeline
  3. Captured before/after view
  4. Visual inspection

02

Vehicle Speed Estimation

Python, PyTorch, YOLO, OpenCV, ByteTrack

View repository

December 2025

Problem

The task is to estimate vehicle speed from video using a known real-world distance, by measuring how long a tracked vehicle takes to travel between two lines in the frame.

Build

The system detects and tracks vehicles, records travel time between configurable virtual lines, and estimates speed from that time and the measured distance. It can retain evidence frames and plate crops.

Evidence

A selected VS13 evaluation reports about 5 percent average error on a side view and about 10 percent on a diagonal view. The repository documents limits from occlusion, camera angle, low light, and low speeds. Those figures depend on the camera setup and are not a system-wide accuracy claim.

Estimation path
  1. Detect
  2. Track
  3. Cross line A
  4. Cross line B
  5. Estimate speed
  1. 03

    Handwritten Digit Recognition

    MNIST classifier with 99.37 percent test accuracy.

    Python, PyTorch

    View repository

    December 2024

  2. 04

    Rotary Autonomous Parking Garage

    Graduation project, graded A+.

    C, C++, Arduino

    August 2021 to January 2022

04 / Skills

Skills and working style

Perception

Computer vision, deep learning, machine learning, object detection, segmentation, tracking, and real-time systems.

Models and data

Model evaluation, data annotation, dataset quality, PyTorch, TensorFlow, Keras, OpenCV, Ultralytics, ONNX, TensorRT, and TensorBoard.

Software

Robotics software, Python, C++, ROS, Docker, Linux, Git, GitHub, and AWS.

Practice

Debugging, refactoring, code review, technical communication, and team collaboration.