Intervia

Multi-agent interview package generator built with CrewAI Flows and exposed via FastAPI.

Given a candidate profile (role, career, experience) and the company's job offer, Intervia produces a complete interview package:

• Market analysis — live web research via Tavily (salary range, responsibilities, in-demand skills, market/industry trends).
• Tailored question bank — questions split by section (soft / hard skills), generated to be answerable on the spot, matching the skills the company asks for.
• Quality review — an evaluator agent judges every question against a fixed rubric and either approves them or sends them back for rework (only the flagged questions are regenerated; approved ones are kept verbatim).

Architecture

intervia/
├── backend-app/
│   ├── app/
│   │   ├── api/        # FastAPI app, routes, JSON error handlers
│   │   ├── schemas/    # Pydantic v2 models (strict JSON at every layer)
│   │   ├── llm/        # LLM provider Strategy: deepseek (default) | anthropic
│   │   ├── agents/     # Agent factories + prompts (one focused agent per stage)
│   │   ├── flows/      # InterviewFlow (@start/@router/@listen) + typed state
│   │   └── core/       # config (pydantic-settings) + exceptions
│   ├── examples/       # curl smoke script + REST Client requests
│   ├── tests/          # mocked end-to-end suite (no live LLMs/Tavily)
│   └── README.md       # setup, endpoints, examples — start here
├── AGENTS.md           # guidance for coding agents working in this repo
└── openspec/           # archived spec-driven design documents

The pipeline is a CrewAI Flow (InterviewFlow) with a deterministic router: analysis agent → routing → question agents (soft and/or hard, in parallel) → evaluator agent loop (approve / rework, capped at 2 rounds) → final package. The models decide the interview focus and the question counts; the code only reads those decisions and assembles the package.

Quick start

cd backend-app
uv sync
cp .env.example .env        # set LLM_PROVIDER, DEEPSEEK_API_KEY, TAVILY_API_KEY
uv run uvicorn app.api.main:app --reload

Then hit POST /api/v1/interviews/analyze — full usage, endpoints and examples are documented in backend-app/README.md.

Design principles

• Models do the work, not the code. Each agent returns a Pydantic schema the next agent consumes; no count math, no grouping helpers, no difficulty labels.
• Consistent evaluation. The evaluator judges against a fixed rubric (relevance, clarity, one-focus, skill coverage, standalone form, immediacy), and questions already approved are never re-judged.
• JSON everywhere, validated by Pydantic v2. Successes return a validated schema; failures return ErrorResponse JSON (error, detail, stage).
• No vendor lock-in. LLM_PROVIDER env var selects the strategy at startup (default deepseek, Claude swappable via config only).

Test

cd backend-app
uv run pytest

The suite never calls live LLMs or Tavily — agents are stubbed at the factory level and the flow is tested end-to-end with mocks.