Nikhil C R
Interview Agent
A stateful interviewer that pauses for each answer, scores it and gives feedback.
Problem
Mock interviews need someone to choose suitable questions, judge answers and give constructive feedback. The agent does that for a chosen role, holding a multi-turn session that waits for a human between steps and adapts the next question to recent performance.
Architecture
Each turn pauses at “await answer”; the API resumes the graph with the candidate’s reply, and the loop continues until the planned number of questions is reached or the bank runs out.
- FastAPIapi.py:84-167
POST /interview/start and POST /interview/{session_id}/submit_answer; the session id is the graph thread id.
- Checkpointerapi.py:64-67
AsyncSqliteSaver persists graph state between requests.
- Question bankdatabase.py
Role-specific questions with topic and difficulty, loaded from MongoDB.
- React + MUI clientInterview.js
Pick a role, answer each question, see feedback alongside.
Key decisions
Human-in-the-loop with interrupt()
The graph pauses with interrupt() at each question and resumes with Command(resume=answer) on the same thread, with state saved in SQLite between HTTP calls.
nodes.py:110-123LLM picks from a bank, and its choice is validated
Given recent turns and the remaining pool, the LLM returns a question id or decides to end; the id is checked against the pool and removed once asked.
llm_helpers.py:97-120One call for analysis and scoring
A single prompt returns relevance, clarity, technical accuracy and sentiment plus a score out of 10, strengths and improvements; required keys are checked.
llm_helpers.py:130-215Explicit terminal states and caps
Status is a typed literal; a failed LLM step terminates cleanly, and the router enforces the planned question count regardless of the LLM.
nodes.py:249-273
In action
Run it
uvicorn agent.api:apicd frontend && npm start