app.py

import gradio as gr
from transformers import pipeline
import requests
import os
import random

# ==============================
# CONFIG
# ==============================

HF_TOKEN = os.getenv("HF_TOKEN")
API_URL = "https://api-inference.huggingface.co/models/mistralai/Mistral-7B-Instruct-v0.2"

headers = {
    "Authorization": f"Bearer {HF_TOKEN}"
}

# ==============================
# Load Whisper (Lightweight)
# ==============================

asr = pipeline(
    "automatic-speech-recognition",
    model="openai/whisper-base"
)

# ==============================
# Question Bank
# ==============================

questions = {
    "Easy": [
        "What is Machine Learning?",
        "Explain supervised learning.",
        "What is overfitting?"
    ],
    "Medium": [
        "Explain bias vs variance tradeoff.",
        "What is gradient descent?",
        "Difference between CNN and RNN?"
    ],
    "Hard": [
        "Explain backpropagation mathematically.",
        "What is attention mechanism?",
        "Explain transformers architecture."
    ]
}

# ==============================
# Generate Question
# ==============================

def start_interview(level):
    return random.choice(questions[level])

# ==============================
# Call LLM via API
# ==============================

def query_llm(prompt):

    payload = {
        "inputs": prompt,
        "parameters": {
            "max_new_tokens": 300,
            "temperature": 0.7
        }
    }

    response = requests.post(API_URL, headers=headers, json=payload)

    if response.status_code == 200:
        return response.json()[0]["generated_text"]
    else:
        return "Error contacting LLM API."

# ==============================
# Evaluate Answer
# ==============================

def evaluate_answer(audio, question):

    if audio is None:
        return "Please record your answer."

    # Speech to Text
    result = asr(audio)
    user_answer = result["text"]

    # Prompt Engineering
    prompt = f"""
You are a strict technical interviewer.

Question:
{question}

Candidate Answer:
{user_answer}

Evaluate and give:

1. Technical Accuracy Score (0-10)
2. Clarity Score (0-10)
3. Depth Score (0-10)
4. Overall Score (0-10)
5. Improvement Suggestions (short and clear)

Be concise and structured.
"""

    feedback = query_llm(prompt)

    return f"""
📝 Transcribed Answer:
{user_answer}

📊 Evaluation:
{feedback}
"""

# ==============================
# UI
# ==============================

with gr.Blocks() as demo:

    gr.Markdown("# 🎤 Smart Interview Simulator (AI Voice Bot)")
    gr.Markdown("Select difficulty → Answer using voice → Get AI feedback")

    level_dropdown = gr.Dropdown(
        ["Easy", "Medium", "Hard"],
        value="Medium",
        label="Select Difficulty"
    )

    question_output = gr.Textbox(label="Interview Question")

    start_button = gr.Button("Start Interview")
    start_button.click(start_interview, inputs=level_dropdown, outputs=question_output)

    audio_input = gr.Audio(
        type="filepath",
        label="Record Your Answer"
    )

    submit_button = gr.Button("Submit Answer")

    result_output = gr.Textbox(label="Evaluation Feedback")

    submit_button.click(
        evaluate_answer,
        inputs=[audio_input, question_output],
        outputs=result_output
    )

demo.launch()