Software Engineer and MSc Student @ University of Windsor
Intro
Welcome to my personal website! Here, you can explore my journey and professional achievements.
If you're interested in a detailed overview of my experience, head over to the Resume tab. For more insight into who I am, including my passions and what drives my work, visit the About Me section. And if you’d like to connect, please feel free to reach out through the Contact tab.
Thanks for visiting, and enjoy exploring!
Resume
Education
M.Sc. Computer Science (AI Stream, Co-op) — University of Windsor (2025–Present)
Coursework: Deep Learning, Natural Language Processing, Software Engineering Topics
B.Comp.Sc. (Honours, Co-op, Software Engineering) — University of Windsor (2019–2024)
GPA: 4.0
Experience
Digital Cloud & Transformations Engineer Intern — Municipality of Lakeshore (Aug–Dec 2025)
Zero Trust networking with Cloudflare WARP, UDP 443/QUIC and NAT timeout diagnostics, MTU/fragmentation analysis,
PowerShell-based monitoring, infrastructure documentation in Confluence and Jira.
Research Assistant — University of Windsor (Jan 2023–Present)
Fine-tuned and evaluated LLMs (BERT, LLaMA, GPT) with feedback-driven optimization, achieving ~20% QA performance gains.
Graduate Assistant — University of Windsor (Jan 2025–Present)
Instructional support for Design & Analysis of Algorithms: dynamic programming, graph algorithms, NP-completeness.
Software Developer Intern — Ground Effects Ltd (Jan 2022–Jan 2023)
Full-stack development with React, Node/Koa, GraphQL; REST APIs in Kotlin/Spring; reporting with SSRS/Power BI;
Python + SQL server migration tooling; automated testing and performance optimization.
Research & Projects
Transformer-Based Automated Interlinear Glossing — Low-resource NLP system achieving 72.5% word-level
accuracy on SIGMORPHON 2023 datasets.
Condensed Variable Semantic Representations — GNN-based code embeddings using VDGs, InfoNCE, and
Deep Graph Infomax (Spearman up to 0.47 on IdBench).
Fine-Tuning LLMs for Question Answering — BERT/LLaMA on SQuADv2.0 with 4-bit quantization
(EM 82.1%, F1 84.1%).
Computer Vision (OpenCV) — Image processing and vision pipelines using OpenCV (see Projects section).
Graduation with Great Distinction (Top 5%) — University of Windsor
Dean’s Renewable Entrance Scholarship
Golden Key International Honour Society (Top 15%)
Competitive Research & Academic Funding — University of Windsor
Work & Research
Research focus: I work on improving sentence representations and self-supervised objectives for modern language models—bridging lightweight sentence encoders with LLMs such as BERT, LLaMA, and GPT-oss. My interests include contrastive learning, better negative mining, layer-wise representation analysis, and evaluation across semantic similarity, retrieval, and downstream NLP tasks.
Research Experience
University of Windsor – Research Assistant
Enhanced textual embeddings in BERT, LLaMA, and GPT-oss using contrastive‐learning objectives and custom loss functions.
Built a Transformer‐based interlinear glossing system, achieving 27–28% higher accuracy on SIGMORPHON 2023 shared‐task datasets.
Developed end‐to‐end pipelines for Question Answering and Sentiment Classification, improving QA accuracy by 15% and sentiment F1 by 10%.
Condensed Variable Semantic Representations
Constructed Variable Dependency Graphs (VDGs) from CodeSearchNet and fused VarCLR embeddings with graph features via a GCN encoder.
Employed InfoNCE contrastive loss and Deep Graph Infomax, achieving Spearman correlations of 0.47 (small split) and 0.45 (medium split) on IdBench.
Introduced a fallback projection for out‐of‐graph variables, retaining over 95% of full‐graph performance.
Hybrid Graph-Based Recommendation System
Developed a GNN + AutoEncoder movie recommender, achieving 90% rating‐prediction accuracy and an RMSE of 0.85.
Implemented adaptive thresholding based on users’ average history, boosting precision by 20% over non‐graph methods.
Work Experience & Development
Municipality of Lakeshore – Digital Cloud & Transformations Engineer Intern
Engineered and monitored secure cloud connectivity for 50+ municipal endpoints by supporting deployment and operation of the Cloudflare WARP client within a Zero Trust network architecture.
Designed and implemented advanced PowerShell-based network monitoring scripts to analyze UDP traffic (including UDP 443 / QUIC), detect NAT timeout behavior, and assess tunnel stability for remote users.
Developed diagnostic tooling to measure MTU path constraints, packet fragmentation, and transport reliability across wired and wireless networks, reducing root-cause analysis time by 30%.
Authored and maintained technical documentation in Confluence and tracked infrastructure issues and remediation tasks in Jira, contributing to standardized and auditable IT operations.
Ground Effects Ltd. – Software Developer Co-op
Built ReactJS frontends and Node.js/Koa/GraphQL backends for client-facing dashboards.
Created paginated SSRS and Power BI reports, delivering actionable insights to project managers.
Developed Python / T-SQL scripts to detect and fix 100+ database anomalies—reducing query times by 15% during a major server migration.
Automated end-to-end testing with Selenium and load-testing with Locust—identifying 15+ critical bugs, boosting throughput by 25%, and cutting page-load times by 20%.
Designed REST APIs in Kotlin/Spring Boot with caching and asynchronous processing, slashing response times by 30%.
University of Windsor – Graduate Assistant
Instruct hands-on Python labs covering data structures, algorithms, and debugging techniques.
Tutor students on algorithmic concepts such as Dynamic Programming, Graph Algorithms, and NP-Completeness, translating theory into practical code.
Collaborate with faculty to design assignments that reinforce algorithmic concepts and coding best practices.
Side Projects & Open-Source Development
Computer Vision project (OpenCV) — experiments and pipelines for vision processing. GitHub repo.
Transformer-based interlinear glossing app for low-resource languages—implemented unsupervised morpheme segmentation and cross-attentive decoding.
Flask-based blogging platform with Dockerized CI/CD (Jenkins), functional tests (Selenium), and load-testing (Locust)—improved test speed by 40% and performance by 15%.
Fine-tuned BERT and LLaMA on SQuAD v2.0 using 4-bit quantization—achieved over 80% Exact Match accuracy while halving training time.
Built a GTK4 + FFmpeg media player with multithreading for buffering, reducing startup latency by 30% and maintaining 60 FPS playback.
About
I’m George Kaceli, a passionate software developer with a Bachelor’s degree in Computer Science from the University of Windsor, graduating with a 4.0 GPA. With extensive experience in both front-end and back-end development, I have worked on various projects, from optimizing database queries to building scalable web applications. I’ve interned at Ground Effects Ltd, where I streamlined workflows for multiple software projects, enhancing efficiency and system reliability through rigorous testing and innovative solutions.
I’m skilled in languages such as Java, Python, and C++, and adept at frameworks like ReactJS, Flask, and Spring Boot. My experience includes fine-tuning machine learning models for question-answering tasks, implementing graph-based recommendation systems, and developing full-stack applications with robust CI/CD pipelines. Whether it's building REST APIs, containerizing applications with Docker, or leveraging cloud technologies, I’m driven to create impactful solutions that optimize performance and user experience.
In my spare time, I enjoy exploring new technologies and working on projects that push my skills in machine learning, data science, and software engineering. I’m excited to continue my journey in tech, solving complex problems and contributing to innovative projects.
Navigate from Start to Target by moving along connected nodes that are semantically closer to the goal.
Each move gives “warmer/colder” feedback, simulating traversal in embedding space.
