Exam MakerExam Maker

Sole engineer: exam generation, the timed session, the grading pipeline, analytics and the bilingual interface.

Preparing for technical interviews usually means generic question banks that ignore the role, the seniority and the stack, and practice tools that show a score without explaining it. AI-generated quizzes often leak their answers to the browser or let the model grade multiple-choice questions it can get wrong.

The candidate sets a target role, an optional job description or focus areas, the level, the format (MCQ, written or mixed), the difficulty curve, the number of questions, the duration and the language. Gemini writes a tailored exam with question weights normalized to exactly 100 points, and a ready gate summarizes the session before the clock starts.

Grading is split by design: multiple-choice answers are scored in code against the sealed key, and written answers are evaluated by Gemini against a structured rubric on five dimensions (correctness, completeness, technical understanding, relevance and clarity), with what was right, what was missing, mistakes and a model answer. The results show the final score, pass status, performance by difficulty and by topic, and an AI debrief with strengths, growth areas, interview readiness and a study plan.

Next.js 16 route handlers call Gemini through @google/genai, every generated exam is validated with Zod, and every grading response is normalized and clamped before use. Answer keys and rubrics never reach the browser before submission: the server seals them with AES-256-GCM and hands the client only the public questions and an encrypted token. Attempts are stored locally in IndexedDB, with no account and no server database. An offline generator keeps the app usable without an API key.

Making a browser exam feel real and fair: a timestamp-based countdown that survives refreshes and background tabs, 5-minute and 1-minute warnings, automatic submission and lock at 00:00, a navigator for answered, unanswered and flagged questions, and a submit dialog that warns about anything left open. Grading free text consistently against a rubric.

The model writes and explains, but never decides a multiple-choice score. Keys are sealed server-side so they cannot be read from the page. History, side-by-side comparison and targeted practice (missed questions or weak areas) turn one exam into a study loop.

Two complete exams were generated, taken and graded with the real model. A Senior Frontend Engineer exam (10 questions, 30 minutes) scored 95/100 with 10 of 10 correct and per-topic scores from 88% to 100%. A Backend Engineer exam (Node.js and PostgreSQL) scored 97/100, with written answers on index design for a 50-million-row table and a resilient job queue scored 26/28 and 34/35. History and Compare show both attempts side by side. Its Vitest suite has 24 tests across 10 files.