考试教练 AI

The problem

Certification candidates often study from static PDFs, disconnected question banks, and generic practice tests that do not adjust to their weak areas. They need realistic questions, clear explanations, progress history, and a way to turn their own material into practice without manually writing every question.

The approach

I built a signed-in practice workspace around Firebase and Vertex AI. The app stores a catalog of certification exams with domain prompts, generates new questions when the Firestore bank does not have enough unused questions, hashes and saves generated questions to avoid duplicates, records taken question IDs per user, and saves every exam result for later review. The dashboard adds recent activity, accuracy trends, study streaks, exam history, CSV export, bookmarks, and AI-generated study topics aligned to official exam domains.

Notable decisions

• Cloudflare Workers deployment with OpenNext. The app uses a custom Vertex AI REST client and Firestore REST adapter because Google SDK internals were not a clean fit for the Workers runtime.
• Question generation is demand-driven. Existing Firestore questions are reused first, then Gemini generates only the missing count, with generated questions stored back into the shared question bank.
• Per-user question history prevents repeat questions by filtering against a userQuestions collection before each new practice session.
• Custom exams support user-defined topics, levels, focus areas, language, public/private visibility, and optional Firebase Storage uploads for source material.
• Stripe subscriptions are modeled as Free, Pro monthly, and Pro annual plans, with webhook handling for checkout, subscription updates, cancellation, and failed payments.
• The study-topics system combines official exam domains, failed-question history, domain weights, and generated recommendations so users can see both weak areas and missing blueprint coverage.

Status

Live at examcoachai.com and actively developed. Recent work includes the Atrium redesign, Cloudflare deployment hardening, Firebase read optimizations, cursor-based exam history pagination, batched exam-detail loading, SEO metadata and structured data, and support for AI help during practice and review.