Muhammad.
Kanvis
Muhammad Abdullah

Muhammad Abdullah

Computer Engineering @ UofT | ML Infrastructure & AI Agentic Systems | Saved $152K+ in industrial ML optimizations
Hire Me

Overview

Computer Engineering student at the University of Toronto with a track record of deploying production-grade ML infrastructure. From saving $152K in manufacturing rejections at Great Lakes Copper to architecting safety-first AI agents with 768-dim semantic memory, I specialize in bridging the gap between complex algorithms and high-performance, real-world applications.
London, ON, Canada

Experience

Jun 2025 – Jun 2026

INTERNSHIP

Machine Learning Engineer (Intern)

Great Lakes Copper
London, ON, Canada
Engineered production ML infrastructure across two live systems - GPU-optimized inference pipelines with lazy model loading deployed on Google Cloud Run, cutting cold-start memory 30%. Architected and developed an ML-powered RAG based incident retrieval platform semantically matching against 100k+ records, achieving 90%+ top recommendations accuracy, and a targeted 40% MTTR reduction. Built an ML Eccentricity Recommendation System optimizing press parameters 16% higher top-quality tube yield, 2% better average eccentricity, $152K saved in lower eccentricity rejections. Owned product discovery for an open-ended AI initiative to reduce production defects - defining scope, success metrics, and a phased roadmap through VP-level stakeholder alignment.
•Engineered production ML infrastructure across two live systems - GPU-optimized inference pipelines with lazy model loading deployed on Google Cloud Run, cutting cold-start memory 30%
•Architected and developed an ML-powered RAG based incident retrieval platform semantically matching against 100k+ records, achieving 90%+ top recommendations accuracy, and a targeted 40% MTTR reduction.
•Built an ML Eccentricity Recommendation System optimizing press parameters 16% higher top-quality tube yield, 2% better average eccentricity, $152K saved in lower eccentricity rejections.
•Owned product discovery for an open-ended AI initiative to reduce production defects - defining scope, success metrics, and a phased roadmap through VP-level stakeholder alignment.

Key Achievements

→Cutting cold-start memory 30%
→Achieving 90%+ top recommendations accuracy
→Targeted 40% MTTR reduction
→16% higher top-quality tube yield
→2% better average eccentricity
→$152K saved in lower eccentricity rejections
MLGPUGoogle Cloud RunRAGSemantic Matching

May 2023 – Aug 2023

INTERNSHIP

AI Intern

Dr. Amir Ahmad, United Arab Emirates University
Abu Dhabi, UAE
Benchmarked 4 anomaly detection algorithms (LOF, KDE, Isolation Forest, OCSVM) on one-class classification and imbalanced datasets, identifying the top performer at 67.4% ROC-AUC across 4 dataset samples. Designed a train-test-validation evaluation framework in Python, analyzing ROC-AUC and precision-recall metrics to characterize each algorithm's strengths and failure modes across diverse data distributions. Built reproducible ML experimentation workflows using Scikit-Learn, PyOD, NumPy, pandas, and matplotlib for preprocessing, training, and results visualization.
•Benchmarked 4 anomaly detection algorithms (LOF, KDE, Isolation Forest, OCSVM) on one-class classification and imbalanced datasets, identifying the top performer at 67.4% ROC-AUC across 4 dataset samples.
•Designed a train-test-validation evaluation framework in Python, analyzing ROC-AUC and precision-recall metrics to characterize each algorithm's strengths and failure modes across diverse data distributions.
•Built reproducible ML experimentation workflows using Scikit-Learn, PyOD, NumPy, pandas, and matplotlib for preprocessing, training, and results visualization.

Key Achievements

→Identifying the top performer at 67.4% ROC-AUC across 4 dataset samples
LOFKDEIsolation ForestOCSVMPythonScikit-LearnPyODNumPypandasmatplotlib

Project Portfolio

ACADEMIC

Present

Urban Planner Mapping Application

Developer

Developed a C++ map application for urban planners using the OSM/Streets Database API, GTK, and EZGL libraries to render maps, features, street names, and POIs. Reduced debugging time by 25% through Git version control and unit testing, and boosted performance by 35% using STL data structures and a multi-threaded Dijkstra algorithm for faster shortest-path computation.
C++OSM APIStreets Database APIGTKEZGLGitSTLDijkstra algorithm
ACADEMIC

Present

Flappy Bird

Programmer

Programmed a Flappy Bird game in C with VGA-based board rendering, FPGA-driven score display, and PS/2 keyboard input for bird control and player switching via an instructions menu. Employed pass-by-reference for large C arrays storing game images, avoiding unnecessary copies to reduce memory overhead and improve rendering performance.
CVGAFPGAPS/2
PERSONAL

Present

Traffic Sign Recognition

Tester/Developer

Designed and executed test cases for an ANN traffic sign classifier functional and regression testing across a 35,000+ image data loader and 40+ classes, resolving class-imbalance and edge-case defects to maintain 85% validation accuracy consistency. Built an automated CI/CD testing environment for the ANN-to-CNN pipeline, collaborating with developers and PMs to accelerate defect identification and resolution.
ANNCNNCI/CD
PERSONAL

Present

Personal-Use AI Agent

Architect

Architected a safety-first macOS AI agent (hexagonal Python core, custom LLM orchestration layer, 12 sandboxed tools, 3 memory modes); grew test coverage 340 -> 599 (+76%) while cutting tool attack surface 21 -> 12 (-43%). Built a hybrid lexical/vector semantic memory system (768-dim Gemini embeddings, LRU caching) with a local privacy-preserving NLP intent classifier, served through a native Swift <-> Python WebSocket interface.
PythonLLMGemini embeddingsSwiftWebSocketsNLP
PERSONAL

Present

Real-Time Sign Language Recognition

Developer

Developed a real-time sign language detection system using TensorFlow/Keras LSTM models, achieving low-latency inference snappy enough for natural signing speeds. Executed test plans for the video preprocessing pipeline frame extraction, normalization, and feature engineering. Evaluated LSTM action-recognition on a live camera feed with OpenCV test environments, measuring spatial-temporal accuracy and logging misclassifications for performance tuning.
TensorFlowKerasLSTMOpenCV

Education

Sep 2022 – Apr 2027

BASc. in Computer Engineering, 2T6+PEY

University of TorontoSpecialization in Computer Engineering

Relevant Coursework

Operating Systems

Vision & Goals

Core Drivers

What are my motivations

Building high-performance AI infrastructure

Solving complex industrial problems with ML

Architecting safety-first agentic systems

Skills & Interests

Python

Git

GitHub Copilot

Cursor

C

Tensorflow

Pandas

REST API

C++

Scikit-Learn

Cloud Run

GCP

Numpy

RAG

Semantic Search

Vector Embeddings

Pytorch

Agile

JavaScript

HTML5

CSS3

SQL

Claude Code

Gemini CLI

FastAPI

React

WebSockets

SQLite

BigQuery

GCS

Docker

CI/CD

JWT

SDLC

MLflow

Codex

Swift

Zustand

Vite

GNNs

Kubernetes

Duo SSO

Optuna

Let's Connect

I'd love to hear from you. Feel free to reach out.

These unlock once you finish onboarding and publish your profile.

More people on Kanvis

Browse more people