da-prep：面向英国学位学徒申请者的 AI 模拟面试与测评练习工具

AI mock interviews, practice tests, an application tracker and guides for UK degree apprenticeship applicants.

Features

• Mock interview (/interview): pick an employer and programme from the researched guides (or paste a job advert) and an interview type (motivation, competency, strengths, commercial awareness, technical, ethics); questions use the firm's process, values and reported questions, and marking rates six skills with three next steps. Each question has its own planned theme so none repeat; text, or timed video style (60s) that is spoken only, with no typing: a microphone picker and test, a live level meter, live captions where the browser supports them, your voice recorded and transcribed, and the transcript shown read-only before marking (video style needs https or http://localhost: browsers block the microphone on network addresses such as http://192.168.x.x); optional read-aloud, text-mode dictation and camera self-view; STAR feedback, score and stronger sample answers.
• Practice tests (/practice): 48 original SJT, numerical, verbal and logical questions with explanations, timers and a review of the ones you missed.
• Assessment replicas (/tests): 19 tests that follow the published format of real employer assessments (SHL Verify Interactive numerical, inductive and deductive; Aon-style statements in short and full length; switch puzzles; Cappfinity-style numerical, verbal and critical reasoning, time-recorded; two inbox job simulations; four situational judgement formats including finance scenarios and SHL work scenarios; two work-style questionnaires) with original questions. Larger pools rotate between attempts. Each shows what is confirmed and what is approximated. Engine in lib/assess/ (item kinds, scoring, adaptive approximation, validation), banks in lib/assess/banks/, catalogue in lib/assess/tests.ts.
• Firm mock processes (/mock): Goldman Sachs, J.P. Morgan, Morgan Stanley, Bank of America, HSBC, PwC, Deloitte, KPMG, EY, Barclays, Rolls-Royce, BAE Systems and Lloyds, stage by stage in the firm's real order, with replica tests, video-style questions, exercises and interviews marked against the firm's own values (lib/mockprocess/). Research behind both is in docs/research/02 and 03.
• CV checker (/cv): paste a CV for school-leaver-specific feedback (score, fixed checklist, section notes, bullet rewrites), optionally checked against an employer or advert. Contact details, links and dates of birth are stripped server-side before the text reaches the AI (redactForCv), the CV is not stored, and it counts as one of the free reviews.
• Statement review (/review): choose the employer and the application question (with its word limit where known); feedback on five criteria, plus what employers say about AI. Stories bank with AI STAR builder (/stories).
• Tracker (/tracker): add a programme from the employer guides with its stages as a checklist and last cycle's dates; closing-date warnings and calendar (.ics) export.
• Employers (/employers): 42 sourced guides with an at-a-glance box, what you'd actually do, why-this-firm talking points and links to matching practice; plus finance firms without a degree route.
• Finance hub (/finance): who hires, a month-by-month application calendar and myths checked against sources.
• Progress (/progress): score trend, STAR coverage, and every past interview with full feedback.
• Sectors (/sectors): what is shared by every degree apprenticeship and what differs for seven sector groups.
• Content: process guide, tips (tests, video interviews, assessment centres), timeline, employers, FAQ.
• Data: backup/restore as JSON and clear local data from the account page.
• Optional accounts + cloud sync (Supabase) and free-tier limits + Pro (Stripe): £9.99 a month.

Run

cp .env.example .env.local   # add OPENAI_API_KEY, or set MOCK_AI=1 to develop without one
npm install
npm run dev                  # http://localhost:3000
npm test                     # vitest (unit, integration, database security with PGlite)
npm run test:e2e             # Playwright journeys on desktop and phone sizes
npm run lint && npx tsc --noEmit && npm run build

Without Supabase env vars the app is local-only (data in the browser). Without Stripe / ENFORCE_LIMITS, everything is free.

MOCK_AI=1 serves canned AI responses so every screen can be developed and tested offline. It is ignored in production builds.

Enabling accounts

1. Create a Supabase project and apply supabase/migrations/*.sql (SQL editor or supabase db push). The migration is safe to re-run.
2. Enable Email (magic link) and, optionally, Google in Auth providers; add your site URL to the redirect list.
3. Set NEXT_PUBLIC_SUPABASE_URL, NEXT_PUBLIC_SUPABASE_ANON_KEY, SUPABASE_SERVICE_ROLE_KEY.

Enabling limits and payments

1. In Stripe, create a Pro product with one recurring price that matches lib/plans.ts (£9.99 a month, GBP). Set STRIPE_SECRET_KEY and STRIPE_PRICE_ID. Checkout refuses to take money if the Stripe price's amount, currency or interval differs from lib/plans.ts.
2. Point a webhook at /api/stripe/webhook for checkout.session.completed, checkout.session.async_payment_succeeded, customer.subscription.updated and customer.subscription.deleted; set STRIPE_WEBHOOK_SECRET. Pro is only granted once a payment is marked paid, and the subscription sets profiles.plan. Enable the customer portal in the Stripe dashboard (Settings, Billing, Customer portal) so "Manage or cancel" works.
3. Limits switch on automatically in production once Supabase is configured (set ENFORCE_LIMITS=false to opt out, or true to force them on in development). All AI routes then require sign-in, share a 150-calls-a-day budget per user (DAILY_AI_CALLS in lib/server/guard.ts) plus a daily cap per expensive route, and free users get 2 interviews, 2 practice tests and 2 reviews in total (FREE_INTERVIEWS, FREE_PRACTICE, FREE_REVIEWS in lib/plans.ts, counted against FREE_PERIOD); Pro has fair-use limits only. Starting an interview or mock process issues a signed practice pass (lib/server/pass.ts) that its follow-up questions, marking and transcription must present, so the free allowance can't be skipped.
4. Production: set UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN so rate limits are shared across serverless instances, and NEXT_PUBLIC_OPERATOR_NAME / NEXT_PUBLIC_CONTACT_EMAIL for the legal pages.

Deploying

Any Node host works (Vercel is the simplest). Set the environment variables above, including NEXT_PUBLIC_SITE_URL. CI (.github/workflows/ci.yml) runs lint, typecheck, tests and a build on every push and pull request.

Layout

• app/ pages and API routes (app/api/*); components/ shared UI; lib/ prompts, schemas, question bank, sector packs, store, backup and calendar helpers; lib/server/ service-role helpers (never import from client code); supabase/migrations/; tests/.
• AI provider: OpenAI Responses API in lib/ai.ts. Two tiers, set by OPENAI_MODEL (marking and feedback, default gpt-6.1-sol) and OPENAI_MODEL_FAST (questions and STAR drafts, default gpt-6-luna), and OPENAI_TRANSCRIBE_MODEL (video-style voice answers, default gpt-transcribe). Requests use store: false. Canned dev responses: lib/mocks.ts.

Before going public

See docs/LAUNCH.md for the launch blockers, set-up steps and verdict, and docs/testing/phase4-report.md for test results. In short: the privacy notice and terms need a lawyer's review; Stripe and Google sign-in are untested against live services; Supabase's built-in email is capped at 2 per hour, so add an SMTP provider; employer data in lib/firms/* and lib/employers.ts needs ongoing re-verification.