BSc (Hons) Computer Science & Artificial Intelligence, First-Class — University of Sussex, 2026 Based in Calgary, AB, Canada · Authorized to work in Canada, no sponsorship required · Open to relocation across Canada

I build and evaluate developer tooling around local LLMs — test generation, automated repair, and mutation-based quality measurement — plus applied ML and secure backend work.

llm-testgen-project · Java, Python, Ollama, JUnit 5, Maven, PIT

End-to-end pipeline that generates JUnit 5 tests from Java source, validates them through Maven compile and execute, repairs failures in an iterative loop, and scores test quality with PIT mutation analysis.

• Diagnosed ANSI/control-character contamination in raw model output; sanitizing it raised pipeline success from 43.3% to 86.7% across 40 Java classes
• Mutation-guided augmentation improved mean mutation scores by 14.6–19.4 percentage points
• 2 x 2 x 2 factorial evaluation across model size (Qwen2.5-Coder 32B / DeepSeek-Coder 6.7B), prompt profile, and repair budget

selenium-regression-suite · Java 17, JUnit 5, Selenium WebDriver, Maven, Docker, GitHub Actions

Browser regression tests against a PHP/MySQL security application I had written — auth, 2FA, password reset, RBAC and upload validation.

• 46 tests, page object model, full suite on every push in headless Chrome in 1m43s, nothing skipped or excluded
• Found 7 defects and fixed 6 with a test guarding each — including two account-enumeration disclosures (login error text, and the reset flow revealing accounts by showing or hiding the security question)
• Found the app could not be reproduced from a clean clone (missing base schema) and made that a CI-verified step
• The README also records the negative results, and a ExpectedConditions.stalenessOf gotcha: it detects staleness by catching StaleElementReferenceException, and Chrome 152 throws a different one — so it fails on the very navigation it is meant to wait for

mdc-data-reference-extraction · Python, XML parsing, rule-based NLP

Extracting data citations from scientific full text and classifying each as primary or secondary use.

• Four-stage pipeline raised dev F1 from 0.057 to 0.639, each stage committed with its own measured delta
• Scored a pre-registered holdout once and reported it unchanged: F1 0.285. Type rules fitted on 122 articles reversed on 23 — an accession-format prior held 95% on dev and 42% on holdout, worse than a constant baseline
• Effective sample size proved to be articles, not mentions (4 articles carried 86 of 153 labels); added article-level bootstrap confidence intervals that mention-level intervals had hidden

Team lead (5 people) · 42nd of 1,282 teams, top 3.3%

applied-machine-learning-projects · PyTorch, scikit-learn, Sentence-BERT

Spam classification comparing TF-IDF against Sentence-BERT embeddings, and facial keypoint regression using ResNet transfer learning.

secure-web-application · PHP, MySQL

Authentication and admin system built to OWASP practices: bcrypt hashing, prepared statements, CSRF tokens, XSS output escaping, 2FA, account lockout, role-based access control, and validated file uploads.

3D-Web · Three.js, GLSL, Blender

Browser-based aircraft showroom with custom GLSL shaders and bloom post-processing; aircraft modelled in Blender.

Live demo

Languages Java · Python · SQL · JavaScript · C# · PHP · HTML/CSS Testing Selenium WebDriver · JUnit 5 · PIT mutation testing · Maven · pytest · page object model AI/ML Local LLM inference (Ollama, Qwen2.5-Coder, DeepSeek-Coder) · prompt engineering · evaluation harnesses · PyTorch · scikit-learn Other Git · GitHub Actions · Docker · Linux · MySQL · Three.js · Unity

Junior Software Engineer · Backend Developer · Test / QA Automation · AI/ML Engineer — in Canada, on site, hybrid, or remote.

Reach me at janbehrem@gmail.com or on LinkedIn.