Projects

Security & study

CertPrep Coach

A Microsoft-style exam simulator with an AI vision layer

· 2 min read

period
2026
role
Solo build
status
In use by my team

Turns practice-question PDFs into an interactive, timed exam that looks and scores like the real Microsoft exam, including case studies, drag-and-drop and hotspot questions. An optional AI pass reads exhibit images once and caches the result, so a whole team can study for free.

I built CertPrep Coach while studying for my Microsoft security certifications. Revising from PDF exports meant scrolling page after page with no way to test myself, and the sites that do offer a simulation mode charge for it. I wanted something that felt like sitting the real exam, so I built it.

You point the app at a folder of PDFs and it parses every question into a local SQLite library. From there you can sit a full exam, a 60-question set (each guaranteed at least one case study and two Yes/No questions) or a ten-question quick test. The interface copies the Microsoft exam: a progress header, an HH:MM:SS countdown, a question navigator with review flags, case studies with their own left-hand navigation, and a score report on Microsoft’s 100–1000 scale with a 700 pass line and a breakdown by skill area.

The hard part was the question types that only exist as pictures, such as drag-and-drop ordering and hotspot dropdowns. An optional AI vision layer looks at each exhibit once, in a single call per question, and rebuilds it as a working interactive question: draggable cards with the right number of answer slots, or dropdowns with the correct answers filled in. It also keeps the “correct answer” image out of sight while you practise and recovers missing answer keys.

Cost was a design constraint from day one. AI tokens are only ever spent during a one-off “Prepare” step, and the results are written to a cache folder that can be committed to Git. One person prepares an exam for roughly one or two US dollars, and everyone else on the team pulls the repository and gets the AI-built questions with no API key at all. Without AI, the app falls back to a manual reveal-the-exhibit flow, so it never depends on a paid service.

Highlights

  • Every Microsoft question type: single, multiple, hotspot, drag-and-drop, case study and lab
  • Microsoft-style 100–1000 scoring with per-skill breakdown
  • AI runs once per exam; results are shared through a Git-committed cache
  • Works behind a corporate TLS-inspection proxy (documented certificate fix)

Built with

  • Python
  • Streamlit
  • PyMuPDF
  • SQLite
  • OpenAI vision

What it taught me

Spend the expensive work once and cache it. Designing around a one-time “Prepare” step kept the tool fast, cheap and shareable.

Source on GitHub