Nouman is available for hire
Hire NoumanNouman Riaz Khan
Verified Expert in Engineering
Machine Learning Developer
Islamabad, Pakistan
Toptal member since August 20, 2019
Bio
For the past five years, Nouman's been working as a machine learning developer—focusing on computer vision apps. His latest projects involved the hands-on development of monocular motion capturing and building Smart Stumps, a cricket product that automates an umpire’s decision making. Nouman also has experience in quantitative analysis as a data analyst working on the American Community Survey and open datasets.
Portfolio
Darvis
Scikit-image, Scikit-learn, TensorFlow, Python, Machine Learning, Deep Learning...
Noerric Technologies
Scikit-image, PyTorch, Scikit-learn, TensorFlow, Python, Machine Learning...
RedBuffer
PyTorch, Scikit-image, Scikit-learn, Python, Machine Learning, Deep Learning
Experience
- Machine Learning - 6 years
- Deep Learning - 4 years
- TensorFlow - 4 years
- Python - 4 years
- Generative Pre-trained Transformers (GPT) - 3 years
- Computer Vision - 3 years
- PyTorch - 3 years
- Natural Language Processing (NLP) - 3 years
Availability
Full-time
Preferred Environment
Python, GitHub, Sublime Text, Ubuntu
The most amazing...
...app I've developed involved monocular motion capturing from a single RGB camera.
Work Experience
Machine Learning Engineer
2019 - PRESENT
Darvis
- Researched and developed computer vision applications in R&D mode.
- Implemented real-time object detection using state-of-the-art research methods in various frameworks (TF, Pytorch, and MxNet).
- Worked on vehicle object tracking using Deep Sort algorithm on dashcam videos.
Technologies: Scikit-image, Scikit-learn, TensorFlow, Python, Machine Learning, Deep Learning, Computer Vision, SQL
Machine Learning Engineer
2018 - 2019
Noerric Technologies
- Developed the "Monocular Motion Capture from Single RGB Camera" API which returns an FBX file of the learned motion/animation from YouTube or any source video.
- Worked on human body part detection using image segmentation techniques on the WSHP dataset. The primary goal was to detect legs and then feet.
- Implemented "PersonLab" for real-time human pose estimation that detected the front-foot landing of a cricket bowler.
- Developed segmentation-based batsman-recognition in cricket videos to classify scenarios on the pitch.
- Developed a face-recognition-based attendance system using a TensorFlow implementation of DeepLab. The API considered the first detection as "IN" and last as "OUT."
Technologies: Scikit-image, PyTorch, Scikit-learn, TensorFlow, Python, Machine Learning, Deep Learning, Generative Pre-trained Transformers (GPT), Natural Language Processing (NLP)
Machine Learning Engineer
2017 - 2018
RedBuffer
- Implemented tree detection in satellite images to assess the shadow penetration on a rooftop of any given house. This was for solar panel placement assessment on the roof.
- Developed SSDH (semi-supervised deep hashing) for finding image similarity during content-based image retrievals (where the hashes were learned for images and image similarities are retrieved by the Hamming distance).
- Implemented Yolo-based medical instrument detection in surgery videos; we trained Yolo v2 on custom datasets.
- Developed a solution that enabled image-to-image translation using Pix2Pix for improving detection masks.
Technologies: PyTorch, Scikit-image, Scikit-learn, Python, Machine Learning, Deep Learning
Data Scientist
2016 - 2017
Bundle
- Implemented text summarization for a query-based news timeline. Given a topic and subtopic query, it returned relevant sentences in a timeline fashion describing events in a subtopic.
- Sorted information that disambiguated the geography in articles for local geography classification where multiple administrative geography levels had the same name.
- Developed a solution for news article classification for the categorization of news articles based on topic modeling and supervised learning.
- Collected data from different UK news sources using crawlers and diffbots.
Technologies: Scikit-learn, Python
Data Analyst
2015 - 2016
PredictifyMe
- Worked as a datasets team member to access, analyze, and explore US open data sets for the first flagship product.
- Analyzed large datasets like the American Community Survey (ACS), real estate, voter data, and OpenStreetMap.
- Conducted predictive analytics of socioeconomic factors, real estate prices, and election votes.
- Assisted the AI team in machine-and-statistical modeling and the development team for dataset knowledge delivery.
- Explored open datasets of different verticals to evaluate the commercial value and suggest product ideas.
Technologies: R
Experience
Monocular Motion Capture
The deliverable here was to develop an API for markerless motion capture. Given a single RGB camera video (from YouTube, etc.), the goal is to learn animation and motion from character and apply this to a 3D character in Unity/Blender.
This was achieved using 2D and 3D human pose estimation, pixel-to-real-world camera coordinates transformation, and applying this to a 3D character in Blender.
This was achieved using 2D and 3D human pose estimation, pixel-to-real-world camera coordinates transformation, and applying this to a 3D character in Blender.
Real-time Human Pose Detection
We implemented the PersonLab paper.
MobileNet was used as a back end for real-time performance and was trained on a COCO dataset for human joint detection.
Joint annotation was used to create a disk and used as a segmentation label for each pair of joints. Every joint class was treated as a separate binary class problem and offsets were predicted using regression.
MobileNet was used as a back end for real-time performance and was trained on a COCO dataset for human joint detection.
Joint annotation was used to create a disk and used as a segmentation label for each pair of joints. Every joint class was treated as a separate binary class problem and offsets were predicted using regression.
Real-time Human Part Detection
The model was trained on a WSHP dataset that had annotations for different body parts; the goal was to work on the legs class.
We used MobileNet as the back-end network and it was trained to use binary classes. The results were good with ~25 FPS.
We used MobileNet as the back-end network and it was trained to use binary classes. The results were good with ~25 FPS.
Face Verification-based Attendance System
The deliverable was an API which receives a cropped-face image and returns matching labels from the learned embeddings. The results were then added to database with timestamp.
VGGFace was fine-tuned with a single image per person and inference was performed using distance metrics.
VGGFace was fine-tuned with a single image per person and inference was performed using distance metrics.
Education
2013 - 2015
Master of Science Degree in Electrical Engineering
International Islamic University Islamabad - Islamabad, Pakistan
Skills
Libraries/APIs
Scikit-learn, TensorFlow, PyTorch, Diffbot
Tools
Scikit-image, GitHub
Languages
Python, SQL
Storage
PostgreSQL
Other
Machine Learning, Computer Vision, Natural Language Processing (NLP), Artificial Intelligence (AI), Data Science, Neural Networks, Deep Learning, Generative Pre-trained Transformers (GPT)
Collaboration That Works
How to Work with Toptal
Toptal matches you directly with global industry experts from our network in hours—not weeks or months.
1
Share your needs
Discuss your requirements and refine your scope in a call with a Toptal domain expert.
2
Choose your talent
Get a short list of expertly matched talent within 24 hours to review, interview, and choose from.
3
Start your risk-free talent trial
Work with your chosen talent on a trial basis for up to two weeks. Pay only if you decide to hire them.
Top talent is in high demand.
Start hiring