app.py

import os
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from datetime import datetime
from transformers import pipeline
from gtts import gTTS
import gradio as gr

LLM_MODEL = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"
ASR_MODEL = "openai/whisper-base.en"

text_gen = pipeline("text-generation", model=LLM_MODEL, max_new_tokens=512)
asr = pipeline("automatic-speech-recognition", model=ASR_MODEL)

def tts_to_mp3(text, fname="feedback_tts.mp3", lang="en"):
    tts = gTTS(text=text, lang=lang)
    tts.save(fname)
    return fname

def make_feedback_visual(scores, labels, out_path="feedback_visual.png"):
    plt.figure(figsize=(4,3), dpi=150)
    bars = plt.bar(labels, scores, color=["#4C78A8", "#F58518", "#54A24B"])
    plt.ylim(0, 100)
    plt.title("Feedback Snapshot")
    for b, s in zip(bars, scores):
        plt.text(b.get_x()+b.get_width()/2, s+2, f"{s:.0f}", ha="center", fontsize=9)
    plt.tight_layout()
    plt.savefig(out_path, format="png")
    plt.close()
    return out_path

def heuristic_scores(feedback_text):
    ft = feedback_text.lower()
    concept = 70 + 15*("concept" in ft) - 10*("misconception" in ft)
    strategy = 65 + 15*("strategy" in ft) + 10*("hint" in ft) - 10*("inefficient" in ft)
    accuracy = 60 + 20*("correct" in ft) - 15*("error" in ft)
    import numpy as np
    return [int(np.clip(x, 0, 100)) for x in [concept, strategy, accuracy]]

def scaffold_prompt(problem, student_text):
    return f"""
You are a pedagogy-aware tutor that gives formative feedback that is timely, specific, and goal-referenced.
Structure the response with:
1) Clarify the goal (feeding up),
2) How it’s going with specific, nonevaluative notes (feeding back),
3) Where to next with scaffolded hints (feeding forward),
4) A brief worked example snippet,
5) 2 self-check questions.

Problem:
{problem}

Student work:
{student_text}

Be concise, supportive, and focus on actionable steps, include concrete hints, and avoid generic praise.
"""

def generate_feedback(problem_text, student_text, audio_file=None):
    transcript = ""
    if audio_file:
        try:
            tr = asr(audio_file)
            transcript = tr.get("text", "").strip()
        except Exception as e:
            transcript = f"[ASR error: {e}]"

    combined_student = (student_text or "").strip()
    if transcript:
        combined_student += f"\n[Voice transcription]\n{transcript}"

    prompt = scaffold_prompt(problem_text, combined_student if combined_student else "(No student work provided)")
    gen = text_gen(prompt, do_sample=True, temperature=0.4, top_p=0.9)["generated_text"]
    feedback = gen[len(prompt):].strip() if gen.startswith(prompt) else gen.strip()

    scores = heuristic_scores(feedback)
    img_path = make_feedback_visual(scores, ["Concept", "Strategy", "Accuracy"])

    audio_out = tts_to_mp3(feedback, fname=f"feedback_{datetime.now().strftime('%H%M%S')}.mp3")
    return feedback, audio_out, img_path

with gr.Blocks() as demo:
    gr.Markdown("## GenAI Feedback Assistant")

    with gr.Tab("Text mode"):
        p_txt = gr.Textbox(label="Problem", lines=3)
        s_txt = gr.Textbox(label="Student Work", lines=6)
        btn = gr.Button("Generate Feedback")
        out_text = gr.Textbox(label="Feedback", lines=12)
        out_audio = gr.Audio(label="Feedback (TTS)", type="filepath")
        out_img = gr.Image(label="Visual Snapshot (PNG)", type="filepath")
        btn.click(generate_feedback, inputs=[p_txt, s_txt, gr.State(None)], outputs=[out_text, out_audio, out_img])

    with gr.Tab("Voice mode"):
        p_txt2 = gr.Textbox(label="Problem", lines=3)
        s_txt2 = gr.Textbox(label="Student Work (optional)", lines=6)
        mic = gr.Audio(sources="microphone", type="filepath", label="Explain thinking (optional)")
        btn2 = gr.Button("Generate Feedback")
        out_text2 = gr.Textbox(label="Feedback", lines=12)
        out_audio2 = gr.Audio(label="Feedback (TTS)", type="filepath")
        out_img2 = gr.Image(label="Visual Snapshot (PNG)", type="filepath")
        btn2.click(generate_feedback, inputs=[p_txt2, s_txt2, mic], outputs=[out_text2, out_audio2, out_img2])

demo.launch()