Hawalay
001An AI-driven lost & found platform. OCR, 512-dimensional embeddings and geo-filtered cosine similarity match reports to items — 94.5% classification accuracy across 13 categories.
Muhammad Bilal
Lahore Machine Learning/AI Engineer
I build computer-vision and applied-ML systems end to end — from dataset collection and training to deployment on a Raspberry Pi, a Linux server, and the software around them.
An AI-driven lost & found platform. OCR, 512-dimensional embeddings and geo-filtered cosine similarity match reports to items — 94.5% classification accuracy across 13 categories.
Real-time waste segregation for a government-backed pilot. A YOLOv26 Nano classifier deployed on a Raspberry Pi 4 and a central Linux server, prepared for rollout at UoS.
CNN, MLP and linear SVM benchmarked on VGG16 transfer-learning features — judged on macro F1 and confusion matrices, not just headline accuracy.
A full-stack booking system: room browsing, availability checks and capacity rules, with server-side validation over a PHP/PDO MySQL schema.
An arcade slicing game in C# — spawn curves, drag physics, collision, lives and state screens. Built to understand game loops properly.
A console catalogue on a self-balancing AVL tree: O(log n) search, insert and delete, with modular file-based persistence.
Archive More experiments on GitHub
I'm a machine learning engineer who cares about the last mile — the part where a model leaves the notebook and has to survive messy data, cheap hardware and real users.
I'm finishing a Computer Science degree at UMT Lahore, with most of my time spent on computer vision. My final-year project, Hawalay, pairs document OCR with 512-dimensional semantic embeddings and MongoDB geo-filtering to match lost items to found ones; the classifier behind it reaches 94.5% test accuracy across 13 categories.
At SyntekAi I built NeuroBin end to end — dataset collection, augmentation, training, evaluation, then deployment onto a Raspberry Pi 4 and a central Linux server for a government-backed pilot. I also write the software around the model: FastAPI services, REST integrations, and the occasional full-stack app. A model is only useful once it ships.
SyntekAi · Lahore
Built NeuroBin, a government-backed smart waste segregation system: a real-time YOLOv26 Nano classifier running on a Raspberry Pi 4 and a central Linux server. Owned the pipeline from dataset collection through to the University of Sargodha pilot.
UMT Lahore
Built the AI layer of an AI-driven lost & found platform: Gemini-powered feature extraction, an OCR pipeline on YOLOv8 and EasyOCR, two-stage matching, and a 13-class InceptionResNetV2 classifier at 94.5% test accuracy.
Freelance · International clients
Built and optimised storefronts, wired payment gateways and REST integrations, and read the funnel instead of guessing — landing page changes aimed at real drop-off metrics.
University of Management and Technology · Lahore
CGPA 3.20. Coursework in machine learning, data structures and algorithms, database systems, statistics for data science and software engineering.
Punjab College
Where the first programs happened, and where the decision to keep going got made.
Funkaar Khana
Marketing and social assets in Photoshop and Illustrator. Still the reason my plots and dashboards are readable.
Hiring for an ML/AI role, or have a problem that needs a model rather than a meeting? I'd like to hear about it.