The Problem
A friend of mine had four months until a competitive government entrance exam — 100 multiple-choice questions, 90 minutes, and a penalty for every wrong answer (-1/3 per incorrect response). Hundreds of pages of legal texts, regulations, and procedures to memorize. Some materials were scanned PDFs with no selectable text.
Her study workflow was entirely manual:
- Read PDFs, highlight key sections by hand
- Write her own practice questions — slow and limited
- No way to test herself at scale or under real conditions
- No visibility into which topics she was weakest on
- No structured review process
She was putting in the hours, but couldn't tell if she was actually making progress. The material was too dense and the exam too punishing to leave it to hope.
She needed a system, not more willpower.
What I Built
No off-the-shelf study app matched her exam's exact format and penalty scoring, so I built her an AI exam prep platform from scratch — the same kind of custom AI build I do for clients, this one as a favour for a friend.
The core idea: upload any study material, and AI generates everything you need to learn it.
Smart Content Ingestion
The platform accepts materials in any format — PDF (including scanned documents via AI vision), Word, spreadsheets, URLs, or plain text. Upload once, organized by topic, and the AI does the rest.
Unlimited AI-Generated Practice Tests
This was the centerpiece. She could select topics and difficulty, and the AI would generate fresh questions every time — in two modes:
- Practice mode — study at your own pace with hints and detailed explanations
- Simulation mode — real exam conditions with timed sessions and penalty scoring matching the actual exam format
The AI also has a rework mode that automatically targets her weakest areas, so every session focuses where it matters most.
Visual Study Tools
Some material is easier to understand visually. The platform generates four types of interactive diagrams — org charts, process flows, mind maps, and comparison tables — using a two-pass AI pipeline where one model generates the structure and another reviews it for accuracy. These aren't static images — they're fully interactive and editable.
Alongside diagrams, the platform generates flashcard decks with a fullscreen study mode and self-rating system, plus study guides and swipeable slide decks for quick review sessions. All from the same uploaded materials.
Progress Tracking That Shows Where to Focus
Instead of guessing, the platform shows exactly which topics need work:
- Per-topic accuracy breakdown with trend tracking over time
- AI-powered analysis that explains why a topic is being missed and suggests study strategies
- Automatic rework tests targeting weak areas
How I Approached It
Two technical decisions shaped the project:
Two AI models instead of one. Gemini 2.5 Pro handles complex generation tasks (diagrams, study guides) where reasoning quality matters. Gemini 2.5 Flash handles high-volume tasks (questions, flashcards) where speed and cost matter. This keeps generation fast and affordable without sacrificing quality.
Reload-safe generation. AI generation can take 10-30 seconds. If the user refreshes mid-generation, the work shouldn't be lost. Every task writes its status to the database and uses polling to track progress — refresh the page, and it picks up right where it left off.
The entire platform runs on Next.js 15 with Neon PostgreSQL — serverless, zero cold starts, deployed on Vercel. Lean stack, fast iteration.
The Result
From the first WhatsApp request ("una IA que me ayude a estudiar y que me haga preguntas tipo test del temario") to a live platform she could log into took one week — March 15 to March 22.
When she sat down to try it, here is what the platform did with her own materials, in a single session:
- 7 PDFs ingested across 8 topics — including scanned documents OCR'd via Gemini's vision capability
- 80 multiple-choice questions generated on demand, matching her exact exam format (4 options, single answer, penalty-aware)
- 8 interactive diagrams generated from the same content — org charts, process flows, mind maps
- Practice and simulation modes running against her real temario
- Everything mobile-optimized, accessible from any device
Her early reaction, by WhatsApp:
"Lo poco que lo he utilizado muy bien — añadí un tema y me hizo las preguntas súper bien para repasar y eso."
— Claudia, two weeks after handoff
The platform did what it was built to do: take any uploaded study material and turn it into an AI-powered exam-prep workflow, without her having to write a single question by hand.
What Worked, What Didn't
Honest read after watching real usage:
The test generator was the standout. AI-generated multiple-choice questions matching the exact exam format — 4 options, single correct, penalty-aware — landed immediately. This is the feature she came back to and the one she liked most. Across the broader user pool (her plus two friends she invited), questions were generated at orders of magnitude faster than writing by hand, and the questions read clearly enough to use without editing.
The diagrams were impressive but not yet sticky. The two-pass AI pipeline (one model drafts, another reviews) produced visually solid org charts and process flows, but real study habits didn't yet route through them — they were a "wow" moment more than a daily-use tool. Diagrams need to plug into the test loop (e.g. "you missed three questions on this topic — here's the diagram of the hierarchy you got wrong") to earn their place.
The note and study-guide generation needs another pass. AI-generated summaries are technically correct but generic. For a discipline where the exact wording of legal articles matters, the next iteration needs to anchor study guides to citations from the source material, not free-form paraphrasing.
Onboarding friction is real. Several users uploaded materials, generated a test or two, and didn't return. That's a product problem, not a generation problem — the platform earns its keep on day 7, not day 1, and the current funnel doesn't make that obvious enough.
Roadmap
Where this is going next:
- Tests → Spaced Repetition loop. Missed questions automatically schedule themselves for review at increasing intervals (SM-2 / FSRS). Test results become the input that drives the next session's content, instead of the user having to decide what to study.
- Citation-anchored study guides. Replace free-form summaries with source-anchored guides — every claim links back to a specific paragraph in the uploaded PDF. Less "AI writing," more "AI navigating."
- Configurable exam profiles. The current platform is hardcoded to one Spanish competitive-exam format. The next version externalizes scoring formulas, time limits, question formats, and language into a "profile" — the same engine ships for university exams, professional certifications, language tests.
- Day-7 retention loop. Onboarding emails + a streak system + a weekly progress digest that closes the gap between "I uploaded my PDFs" and "I have a study habit."
- Diagrams as feedback, not standalone artifacts. Diagrams generated in response to wrong answers, anchored to the specific topic the user is missing.
Same engine, any exam.
Facing a similar challenge? I build custom AI solutions that save time and eliminate manual work. Let's talk about what I can build for you.
Have a similar challenge?
I build custom AI solutions that save time and eliminate manual work. Let's talk about what I can build for you.
Book a Call