🚀 InterviewAlly

Your Personal AI Interview Coach. InterviewAlly is a local, privacy-first application that ingests interview recordings (video/audio), analyzes them using RAG (Retrieval-Augmented Generation), and provides brutal, actionable feedback to help candidates get hired.

Unlike generic chatbots, InterviewAlly uses Speaker Diarization to distinguish between the interviewer and candidate, and employs a specialized RAG pipeline to rewrite weak answers using the STAR method based on the candidate's actual experience.

---

⚡ Key Features

• 🎧 Multimodal Ingestion: Upload .mp4, .mov, .mp3, or .wav files. The system uses Faster-Whisper for transcription and Pyannote.audio for Speaker Diarization (who said what).
• 🧠 RAG-Powered Analysis:
  • Indexes transcripts into ChromaDB using Nomic Embeddings (nomic-embed-text-v1.5).
  • Preserves temporal metadata (timestamps) for precise context retrieval.
• 📊 Automated Scoring: Uses Llama 3.2 (via Ollama) to grade Technical Depth, Communication, and Cultural Fit on a 1-10 scale.
• ✨ Magic Rewriter: Identifies weak answers and rewrites them into "Gold Standard" responses using the STAR Method (Situation, Task, Action, Result), grounded strictly in the transcript data (no hallucinations).
• 💬 Chat with your Interview: A Q&A interface to ask specific questions like "Did I sound nervous?" or "How can I improve my explanation of SQL joins?"

---

🛠️ Tech Stack

• LLM & Orchestration: LangChain, Ollama (Llama 3.2), Pydantic (Structured Output).
• Speech Processing: Faster-Whisper, Pyannote.audio.
• Vector Database: ChromaDB (Local persistence).
• Backend: FastAPI, SQLite (Job management), BackgroundTasks.
• Frontend: Streamlit.

---

🏗️ Architecture

1. Ingestion: Video is uploaded -> Audio extracted -> Transcribed (Whisper) -> Speakers Identified (Pyannote).
2. Indexing: Transcript is split into semantic chunks (grouped by speaker turns & time) -> Embedded -> Stored in ChromaDB.
3. Analysis: Background task runs a "Coach" chain to generate a JSON report saved to SQLite.
4. Interaction: User views scores/feedback in Streamlit and triggers RAG-based rewrites or chat.

---

🚀 Setup & Installation

Prerequisites

• Python 3.10+
• Ollama installed and running.
• FFmpeg installed (required for audio processing).

1. Clone the Repository

git clone https://github.com/nkhanna94/InterviewAlly.git
cd InterviewAlly

2. Install Dependencies

pip install -r requirements.txt

3. Setup Models

Pull the LLM:

ollama pull llama3.2:latest

Environment Variables: Create a .env file in the root directory:

# Required for Pyannote Speaker Diarization
HUGGINGFACEHUB_API_TOKEN=your_huggingface_token_here

4. Run the Application

You need to run the Backend and Frontend in separate terminals.

Terminal 1: Backend (FastAPI)

uvicorn backend.main:app --reload

Terminal 2: Frontend (Streamlit)

streamlit run frontend/app.py

---

📂 Project Structure

InterviewAlly/
├── backend/
|   ├── chroma_db/           # Local Vector Store
│   ├── main.py          # FastAPI endpoints & background tasks
│   ├── brain.py         # LLM logic, Chains, & Structured Output
│   ├── rag.py           # ChromaDB setup & Custom Chunking logic
│   ├── transcripts.py   # Whisper + Pyannote pipeline
│   └── jobs.db          # SQLite database for job status
├── frontend/
│   └── app.py           # Streamlit Dashboard
└── requirements.txt

---

🔮 Future Improvements

• LangGraph Integration: Move from linear Chains to a stateful Agent for iterative critique/refinement of answers.
• SQL Agent: Allow users to query their history (e.g., "Show me all interviews where I failed System Design").
• Cloud Deployment: Dockerize the application for AWS/GCP deployment.

---