src/streamlit_app.py

import streamlit as st
from transformers import AutoTokenizer, AutoModelForCausalLM, pipeline
import torch
from PIL import Image
import numpy as np
from gtts import gTTS
import tempfile
import re

# ----------------------------
# MODELS
# ----------------------------
MODEL_NAME = "TinyLlama/TinyLlama-1.1B-Chat-v1.0"

@st.cache_resource
def load_llm():
    tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
    model = AutoModelForCausalLM.from_pretrained(
        MODEL_NAME,
        torch_dtype=torch.float32,
        device_map="cpu"
    )
    return tokenizer, model

@st.cache_resource
def load_asr():
    return pipeline("automatic-speech-recognition", model="openai/whisper-tiny")

tokenizer, model = load_llm()
asr = load_asr()

# ----------------------------
# MULTILINGUAL DETECTION
# ----------------------------
LANG_WORDS = {
    "en": ["one", "two", "three", "four", "five"],
    "fr": ["un", "deux", "trois", "quatre", "cinq"],
    "sw": ["moja", "mbili", "tatu", "nne", "tano"],
    "kin": ["imwe", "ebyiri", "eshatu", "enye", "eshanu"]
}

def detect_mixed_language(text):
    text = text.lower()
    scores = {lang: 0 for lang in LANG_WORDS}

    for lang, words in LANG_WORDS.items():
        for w in words:
            if w in text:
                scores[lang] += 1

    dominant = max(scores, key=scores.get)

    # detect mix
    active_langs = [l for l, s in scores.items() if s > 0]

    if len(active_langs) > 1:
        return dominant, active_langs
    else:
        return dominant, [dominant]

# ----------------------------
# PROMPT ENGINEERING
# ----------------------------
def build_prompt(user_input, dominant_lang, langs_used):
    if dominant_lang == "fr":
        base = "Tu es un tuteur de mathématiques pour enfants. Explique simplement."
    elif dominant_lang == "sw":
        base = "Wewe ni mwalimu wa hesabu kwa watoto. Eleza kwa urahisi."
    elif dominant_lang == "kin":
        base = "Uri umwarimu w'imibare ku bana. Sobanura neza."
    else:
        base = "You are a friendly math tutor for kids. Explain step by step."

    # Handle code-switch
    if len(langs_used) > 1:
        base += " The child used mixed languages. Keep explanation in main language but reuse number words from other language."

    return f"{base}\nUser: {user_input}\nAssistant:"

# ----------------------------
# GENERATION
# ----------------------------
def generate(prompt):
    inputs = tokenizer(prompt, return_tensors="pt")
    output = model.generate(
        **inputs,
        max_new_tokens=80,
        temperature=0.7,
        do_sample=True
    )
    return tokenizer.decode(output[0], skip_special_tokens=True)

# ----------------------------
# TTS (MULTILINGUAL)
# ----------------------------
def speak(text, lang="en"):
    lang_map = {
        "en": "en",
        "fr": "fr",
        "sw": "sw",
        "kin": "en"  # fallback
    }

    tts = gTTS(text=text, lang=lang_map.get(lang, "en"))

    with tempfile.NamedTemporaryFile(delete=False, suffix=".mp3") as fp:
        tts.save(fp.name)
        return fp.name

# ----------------------------
# VISUAL COUNTING (Baseline)
# ----------------------------
def count_objects(image):
    img = np.array(image.convert("L"))
    binary = img > 128
    count = int(binary.sum() / 400)
    return max(1, count)

# ----------------------------
# UI
# ----------------------------
st.set_page_config(layout="wide")
st.title("🧠🌍 Multilingual AI Math Tutor")

col1, col2 = st.columns(2)

# ----------------------------
# LEFT PANEL
# ----------------------------
with col1:
    st.header("👧 Student Interaction")

    mode = st.radio("Mode", ["Text", "Voice", "Image"])

    # -------- TEXT --------
    if mode == "Text":
        user_input = st.text_input("Ask or answer:")

        if user_input:
            dominant, langs = detect_mixed_language(user_input)

            prompt = build_prompt(user_input, dominant, langs)
            response = generate(prompt)

            st.write("### 📘 Answer")
            st.write(response)

            st.write(f"🌍 Dominant: {dominant} | Mixed: {langs}")

            if st.button("🔊 Speak"):
                audio = speak(response, dominant)
                st.audio(audio)

    # -------- VOICE --------
    elif mode == "Voice":
        audio_file = st.file_uploader("Upload voice (.wav)", type=["wav", "mp3"])

        if audio_file:
            result = asr(audio_file)
            text = result["text"]

            st.write(f"🗣️ Detected: {text}")

            dominant, langs = detect_mixed_language(text)

            prompt = build_prompt(text, dominant, langs)
            response = generate(prompt)

            st.write("### 🎧 Response")
            st.write(response)

            audio = speak(response, dominant)
            st.audio(audio)

    # -------- IMAGE --------
    elif mode == "Image":
        uploaded = st.file_uploader("Upload image", type=["png", "jpg"])

        if uploaded:
            image = Image.open(uploaded)
            st.image(image)

            count = count_objects(image)

            st.write(f"### 🧮 I see about {count} objects")

            explanation = f"There are {count} objects. Let's count together."

            audio = speak(explanation)
            st.audio(audio)

# ----------------------------
# RIGHT PANEL (DASHBOARD)
# ----------------------------
with col2:
    st.header("📊 Learning Dashboard")

    st.metric("Questions", 15)
    st.metric("Accuracy", "80%")
    st.metric("Level", "Improving")

    st.subheader("📈 Skill Progress")
    st.progress(0.8)

    st.subheader("🌍 Language System")
    st.write("✔ English / French / Swahili / Kinyarwanda")
    st.write("✔ Code-switch detection")

    st.subheader("⚡ Features")
    st.write("✔ Voice (Whisper)")
    st.write("✔ Visual counting")
    st.write("✔ Multimodal learning")