Interview Prep AI

An adaptive AI interview coach for DSA preparation. Unlike static roadmap tools, it personalizes recommendations, difficulty, and pacing based on your actual performance — pattern mastery tracking, weakness detection, and AI-evaluated approach submissions.

Phase 1 status: MVP backend + minimal frontend complete. Chat (SSE), attempt evaluation, weakness detection, roadmap, and tracker pages all wired end-to-end. See plan.md for the implementation log.

Tech Stack

Layer	Choice
Frontend	React 19 + Vite, Tailwind CSS, Zustand, React Router v6
Backend	Express.js (Node, TypeScript)
Database	PostgreSQL + Prisma ORM
Auth	Clerk (@clerk/clerk-react + @clerk/express)
AI	OpenRouter (single gateway → Claude Sonnet, GPT-4 Turbo, GPT-4o-mini)

All model calls go through one abstraction in server/src/lib/ai/client.ts so swapping providers is a string change.

Project Structure

client/                       React + Vite SPA
  src/
    pages/                    RoadmapPage, TrackerPage, ...
    hooks/useChat.ts          SSE chat hook
    store/userStore.ts        Zustand + Clerk bridge
    lib/api.ts                Typed fetch client
server/                       Express backend
  src/
    routes/                   users, chat, attempts, progress, roadmap, weakness
    lib/
      ai/                     OpenRouter client + versioned prompt files
      db/queries/             Reusable Prisma helpers
      weakness/detect.ts      Passive weakness detection
      recommendation/         Today's plan builder
  prisma/
    schema.prisma             11 normalized tables
    seed.ts                   Idempotent topic/pattern/problem seed
data/                         Human-editable seed JSON
  topics.json                 10 topics
  patterns.json               8 DSA patterns
  problems.json               54 curated problems
plan.md                       Step-by-step implementation log
CLAUDE.md                     Product spec + architecture decisions

Getting Started

Prerequisites

• Node.js 20+
• PostgreSQL 14+ with a database you can write to
• A Clerk project (for auth)
• An OpenRouter API key (for AI calls)

1. Install

npm install

This installs both workspaces (client and server) via npm workspaces.

2. Configure environment

Copy .env.example into both server/.env and client/.env and fill in the values relevant to each side. The example file documents what goes where.

Without CLERK_SECRET_KEY + CLERK_PUBLISHABLE_KEY the server still boots, but every protected route returns 401. This is intentional for smoke tests.

3. Database

cd server
npx prisma migrate dev --name init
cd ..
npm run seed              # loads 10 topics, 8 patterns, 54 problems

To wipe and reseed:

npm run db:reset

4. Run

npm run dev

This launches both servers concurrently:

• Client — http://localhost:5173 (Vite, proxies /api/* to the server)
• Server — http://localhost:4000

Sanity check: curl http://localhost:4000/health → {"status":"ok"}.

Scripts

Command	What it does
npm run dev	Start client + server concurrently
npm run dev:client	Vite only
npm run dev:server	Express only (ts-node-dev --respawn)
npm run build	Type-check + bundle both workspaces
npm run type-check	tsc --noEmit for both
npm run seed	Run the Prisma seed script
npm run db:reset	prisma migrate reset --force && seed

Architecture Notes

• Single AI client. Every model call goes through streamChat / evaluateApproach / generateRoadmap in server/src/lib/ai/client.ts, which routes to OpenRouter and falls back automatically per use case.
• Versioned prompts. Stored as .md files under server/src/lib/ai/prompts/ with frontmatter (<!-- version: 1.0 | updated: ... -->). Loaded via a tiny templating helper, never inlined.
• Tracker is home base. Every write path (POST /api/attempts) updates TopicProgress + PatternMastery in the same transaction and fires passive weakness detection without blocking the response.
• Pattern from AI, not DB. Mastery is updated against the pattern the user actually used (per the AI evaluation), not the canonical pattern tagged on the problem.
• SSE for chat. All chat responses stream token-by-token; the user message is persisted before streaming starts so history survives mid-stream disconnects.

See CLAUDE.md for the full product vision and architecture decisions.

Phase Roadmap

Phase	Features
1 ✅	Scaffold, Clerk auth, AI chat (SSE), AI approach evaluation, static roadmap, tracker, weakness detection
2	Adaptive roadmap (topic graph), pgvector chat memory, recommendation engine
3	Readiness score, PostHog analytics, mock interview mode
4	Voice interviews, exportable reports

License

MIT — see LICENSE.