Back to projects

Job AI Assistant

Upload a CV and job description, get a match score and skill gaps, then take a real-time voice mock interview.

SWE (AI/MLOps)LLM AgentsRAGVoice AILangChainNext.js

A 3-agent pipeline orchestrated by a supervisor state machine: Agent 1 parses a CV and job description with Groq/Llama 3.1 and produces a validated match score, strengths, and gaps; Agent 2 generates behavioral, technical, and role-specific interview questions, calibrating difficulty from those gaps and retrieving similar past questions from a ChromaDB/LlamaIndex RAG store; Agent 3 conducts a real-time voice interview over LiveKit, transcribing answers with a local Whisper model, evaluating them with Groq, and speaking follow-ups and feedback back with Coqui TTS. Every agent inherits a shared self-healing loop, Pydantic-validated output, and on failure a self-critique-and-retry cycle (up to 3 attempts) before escalating. A FastAPI backend with WebSocket push updates drives a Next.js frontend through upload, live analysis, question review, and the voice interview itself, with MLflow tracking quality score, retry count, latency, and token usage per agent run.

What I built

  • Built a 3-agent pipeline (CV/JD analyzer, question generator, voice interviewer) behind a Supervisor MCP server that auto-advances agents through an explicit state machine.
  • Gave every agent a shared self-healing loop: Pydantic v2 schema validation on every LLM output, with a self-critique-and-retry cycle (max 3 attempts) on failure before escalating.
  • Built a RAG pipeline (LlamaIndex chunking, SentenceTransformer embeddings, ChromaDB cosine-similarity search) to retrieve similar past interview questions and inject them into the question-generation prompt.
  • Built a real-time voice interview agent on LiveKit WebRTC, using a local Whisper model for speech-to-text and Coqui TTS for spoken feedback, with automatic follow-up questions when an answer scores 5 or below.
  • Exposed the pipeline through a FastAPI backend with WebSocket push updates (state-change-only broadcasts, not polling) and a Next.js frontend that progressively reveals the match score, questions, and interview UI as each agent completes.
  • Used Pydantic model_validator to auto-compute derived fields (average score, question counts, pass/fail) from LLM output rather than trusting the model to self-report them.
  • Tracked quality score, retry count, latency, and token usage per agent run through MLflow.

Tech stack

SWE (AI/MLOps)LLM AgentsRAGVoice AILangChainNext.js
Back to all projects