Feras Mahmood · software developer | Researcher
I build dependable AI systems, from data pipeline to inference.
I engineer agentic workflows, machine-learning infrastructure, and data platforms designed for AI agents and real users for compliance heavy products.
My Profile
I am a final-year Computer Science student at Sheridan College specializing in GenAI and ML development. Currently, I work as an AI Software Developer and Research Assistant, leading the deployment of AI infra and machine learning services that support 10,000-15,000 active users. I am particularly driven by engineering data pipelines that transform manual verification into automated, scalable workflows.
Outside work, I play soccer and love experimenting with new recipes in the kitchen.
Work Experience
Data AnalystParadigm Electronics Inc.Mississauga, ONMay 2024 – Aug 2025
Top Projects
Completed
HR / Talent Acquisition App
Smart Recruitment Platform that streamlines the hiring process by automatically analyzing resumes and verifying candidate backgrounds.
The system employs Autonomous AI Agents that act like digital recruiters. They use web browsing tools to verify candidate claims, cross-reference experience on external sites, and conduct research to present a validated report to hiring managers.
Built for scale and reliability, the system features a full observability stack with Prometheus (metrics),Grafana (dashboards), and Tempo (distributed tracing). The infrastructure is fully containerized withDocker (featuring "baked-in" models) and powered by MongoDB for event-driven persistence.
Deployed on: Microsoft Azure
Key Technical Highlights
- Micro-Service Architecture: Decoupled agents using Model Context Protocol (MCP).
- Production Observability: Full metric tracking with Prometheus & Grafana.
- Containerization: Optimized Docker builds with pre-loaded models for performance.
- Event-Driven Design: Asynchronous pipelines triggered by database events.
Next.jsReactTypeScriptLangChainLangGraphLlama 3LlamaParsePuppeteerMongoDBAzure Container AppsGitHub ActionsOpenTelemetry
Completed
OptiMulti-Video
High-Performance Multimodal Attention with Custom CUDA Kernels. This project demonstrates a vertical slice of a high-performance Multimodal AI system.
It features a Custom CUDA Kernel written in CUDA C++ for low-latency fusion of video and text embeddings, integrated into a compact Video-Text Transformer architecture.
The training pipeline utilizes Fully Sharded Data Parallel (FSDP) to distribute workloads across dual T4 GPUs, optimizing for both performance and resource accessibility.
Tested on: Google Colab / Kaggle (Dual T4 GPUs)
Key Technical Highlights
- Custom CUDA Kernel: Fused "Normalize & Project" for low latency.
- Distributed Training: FSDP implementation for multi-GPU setups.
- Multimodal Fusion: Video-Text Transformer architecture.
CUDA C++PyTorchFSDPPythonGoogle ColabTransformer
Completed
Sports Analytics App
Production-Grade Computer Vision Application developed to analyze cricket shot mechanics in real-time. This mobile platform utilizes YOLOv8 for high-speed object detection to track ball trajectories and RandomForest classifiers to provide instant, data-driven shot recommendations and corrective feedback to players.
The system features a cross-platform mobile app built with Flutter, offering a seamless user experience for recording execution and receiving detailed coaching insights. Inference is decoupled via a serverless architecture on Google Cloud Platform (Cloud Run), ensuring cost-effective scaling to zero when idle, while Pub/Sub handles asynchronous event messaging for smooth analysis workflows.
Deployed on: Google Play Store (Internal Testing) & GCP
Achievement
- First Place for Capstone projects in data analytics stream.
Impact Metrics
- Deployed to production with ~10 active users.
- Real-time CV inference via serverless architecture.
- Cost-optimized ML inference using Cloud Run.
Key Contributions
- Designed end-to-end CV pipeline using YOLOv8 to detect ball trajectories.
- Deployed scalable inference using Cloud Run.
PythonOpenCVYOLOv8RandomForestFlutterGoogle Cloud PlatformCloud RunPub/Sub
Writing
Infrastructure ·
Managed LLM APIs vs. Your Own vLLM Server
Managed APIs and self-hosted inference solve different workload shapes. The useful comparison starts with utilization, repeated context, latency, and operational control, not ideology.
Certifications
In Progress
Databricks Certified Data Engineer Associate
Databricks
Endorsements
Tom Khirdaji
Information Technology Manager at Paradigm | Anthem | Martin Logan Electronics
“I had the pleasure of supervising Feras Mahmood during his 16-month internship as a Data Analyst at Paradigm. Over this time, Feras consistently demonstrated strong technical skills, curiosity, and a commitment to learning. Feras worked on a wide range of projects, delivering KPI dashboards, writing SQL queries, and integrating data through RESTful APIs. He also contributed to automation efforts using PowerShell scripts, which streamlined some of our internal workflows. Beyond analytics, Feras showed initiative by building several web applications using JavaScript, Node.js, and React, expanding his impact beyond traditional data analysis. What stood out about Feras was his ability to quickly grasp new tools and technologies, and his willingness to take ownership of tasks. His contributions helped our team improve data visibility and operational efficiency. I am confident that his blend of analytical and development skills will serve him well in his future career, and I would gladly recommend him for any data or technology role he pursues.”
Frank Pugliano
Manager at Paradigm Electronics
“Feras played a pivotal role in developing high-impact internal web applications for our team, along with dashboards that gave us significantly greater visibility into client documents and key financial metrics. These tools made a real difference in how our team operated day-to-day.”
Volunteering
Workshop Team MemberGoogle Developer Group SheridanMay 2025 – Present
- Supported and organized the design and delivery of community tech workshops, helping 100+ attendees gain hands-on experience with Google technologies.
Co-OrganizerDatathon (Google Developer Group)Oct 2025
- Co-organized a 300+ people hackathon 'Datathon' with Google Developer Group, managing logistics and mentor support.
- Event Link ↗
Contact
Feel free to reach out to me via email or LinkedIn.
technocratz979@gmail.com →