What I've built and completed

This portfolio site

A small static site built and shipped as part of FlyRank's Week 5 assignment — the point was getting something real and live rather than a polished mockup. Built with plain HTML and CSS, no framework, so every part of it is something I can explain line by line.

• Repo: github.com/itretzehra-max/flyrank-project
• Deployed via GitHub Actions to GitHub Pages

LinkedIn profile rebuild

Ran my LinkedIn through FlyRank's audit tool (started at 47/100), then went through it flag by flag: rewrote role descriptions that had zero content, added concrete scope to claims that were previously vague, tied my skills to the education and experience entries where I actually used them, and removed a skill (SQL) I hadn't learned yet instead of leaving it unbacked.

AI Risk Register Generator

The problem: Manually classifying AI systems under the EU AI Act's risk tiers takes too long for busy compliance teams, and it's easy to miss an obligation buried in the regulation's text.

What I did: Built a Python tool (originally wrapped in a Streamlit web app) that asks structured yes/no questions about how an AI system is used — does it fall under a prohibited use, does it operate in a high-risk domain like employment or credit, does it interact directly with users in a way that could deceive them — and maps the answers to the correct EU AI Act risk tier plus the specific mitigations that tier requires.

What I found (example run): Classifying a sample AI-powered resume-screening tool, the system flagged it as High Risk (used in the employment domain) and automatically listed the four required mitigations: a conformity assessment, technical documentation and a risk management system, human oversight with override capability, and activity logging for auditability.

What came of it: A 2-minute interactive session that replaces a manual read-through of the regulation with a structured decision tree — someone can classify a system and get the exact list of compliance obligations, not just a vague "this might be risky."

Code: risk_register.py

AI System Compliance Case Study

The problem: Organizations deploy AI models on decision data (like loan approvals) without checking whether outcomes differ unfairly across demographic groups — creating legal exposure and hidden bias.

What I did: Built a Python/Pandas script that audits a sample loan-approval dataset (30 applicants, split evenly by gender) for bias. It calculates approval rates by group, applies the "four-fifths rule" (a standard regulatory fairness threshold), and cross-checks whether a legitimate factor like credit score actually explains any gap.

What I found: On the sample dataset, female applicants had a 40% approval rate versus 73.3% for male applicants — a 33-point gap. The four-fifths rule flagged this immediately (selection rate ratio of 0.55, well below the 0.80 threshold). But the average credit score gap between groups was only 19 points — too small to explain a 33-point approval gap on its own, meaning the disparity likely isn't fully justified by credit risk alone.

What came of it: A reusable audit script that gives compliance teams a concrete, quantified flag — not just "check for bias" but an actual pass/fail signal plus the reasoning behind it, so a human reviewer knows exactly what to investigate next.

Code: bias_audit.py · sample_loan_data.csv

Live Contact Form

The problem: The portfolio site was static — visitors had no way to actually reach out beyond a bare mailto link, so there was no real lead-capture point on the site.

What I did: Wired up a working contact form using Formspree, so submissions on the Contact page get sent directly to my inbox without needing a backend of my own.

What came of it: A functioning dynamic feature on an otherwise static site, plus a plain-words map of the data flow: visitor fills form → Formspree receives the POST → email lands in my inbox.