AI Interview

Text-first resume assessment and interview practice application for technical job seekers.

The product lets users upload a resume, optionally provide a job description, receive a structured resume strength assessment, generate tailored interview questions, answer them, and get feedback.

Current State

This repository currently contains the Spring Boot backend, the first local infrastructure setup, and a Next.js frontend prototype. The backend prototype is no-auth and single-resume for now.

• Spring Boot API backend.
• Next.js frontend.
• PostgreSQL with pgvector.
• Redis.
• LocalStack S3-compatible object storage.

See docs/project-design.md for the proposed architecture, domain model, API surface, repository structure, and implementation milestones.

Local Infrastructure

Create local environment config:

cp .env.example .env

Set DATABASE_URL, DATABASE_USERNAME, and DATABASE_PASSWORD in .env to match your existing PostgreSQL/pgvector instance. The local default database is interview_guide on localhost:5432.

Start the local services:

docker compose up -d

This starts:

• Redis on localhost:6380 by default when managed by Compose
• LocalStack S3 on localhost:4566

The backend defaults to PostgreSQL on localhost:5432 because this workspace already has pgvector there. If you need Compose to manage a separate PostgreSQL instance later, run:

docker compose --profile managed-postgres up -d

The managed PostgreSQL service binds to localhost:55432 by default, or to POSTGRES_PORT if set.

The backend defaults to Redis on localhost:6379. The bundled Compose Redis service publishes container port 6379 to localhost:6380 by default, so set REDIS_PORT=6380 only if you are using that Compose-managed Redis instance. If your Docker Redis is already mapped to host port 6379, leave REDIS_PORT=6379.

Redis is used for operational guardrails only: per-client limits on expensive upload/AI requests and short-lived duplicate in-flight locks around Gemini calls. PostgreSQL remains the durable source of truth.

Mutation endpoints also support an optional Idempotency-Key header. When present, Redis stores the successful response for the same client, endpoint, and request fingerprint for REDIS_IDEMPOTENCY_TTL_SECONDS seconds. Retrying the same request with the same key returns the cached response; reusing the key with a different payload returns 409.

The backend stores original uploaded resume files in S3-compatible storage and defaults to LocalStack:

S3_ENDPOINT=http://localhost:4566
S3_REGION=us-east-1
S3_BUCKET=ai-interview
S3_ACCESS_KEY=test
S3_SECRET_KEY=test

Application-owned persistence tables are created under the ai_interview_app schema to avoid collisions with existing local tables in public.

Gemini calls use the Gemini Developer API key from local .env. To enable Gemini chat and embeddings for development, update .env:

GEMINI_API_KEY=your-api-key
AI_CHAT_MODEL=google-genai
AI_EMBEDDING_MODEL=google-genai
GEMINI_CHAT_MODEL=gemini-2.5-flash
GEMINI_REQUEST_TIMEOUT_SECONDS=90
GEMINI_MAX_OUTPUT_TOKENS=2048
GEMINI_THINKING_BUDGET=0
GEMINI_EMBEDDING_MODEL=gemini-embedding-001
GEMINI_EMBEDDING_DIMENSIONS=1024
RAG_EMBEDDING_DIMENSIONS=1024

Keep .env local and untracked. Do not put real API keys in .env.example.

First Implementation Milestones

1. Set up project foundation and local infrastructure.
2. Add resume upload and text extraction.
3. Add resume assessment with optional job description matching.
4. Add interview question generation and answer feedback.
5. Harden with rate limits, tests, observability, and schema validation.

Backend API

The first no-auth API persists uploaded resumes, extracted chunks, generated assessments, interview sessions, questions, and answer feedback for a single local user:

• POST /api/resumes with multipart field file
• GET /api/resumes/current
• POST /api/assessments
• POST /api/interview/questions
• POST /api/interview/feedback

Supported upload formats: PDF, DOC, DOCX, TXT, and Markdown.

The AI endpoints call Gemini and return typed JSON for the frontend. If Gemini is unavailable or the configured key/model has no quota, the frontend falls back to local draft scoring so the workflow remains usable during development.