Deploy AIRA v2 - Qwen vision + voice + heat safety

README.md

@@ -1,13 +1,97 @@
 ---
-title: Aira V2
-emoji: 🐠
-colorFrom: pink
-colorTo: purple
+title: AIRA v2 — Arid-climate Intelligent Risk Assistant
+emoji: 🌡️
+colorFrom: yellow
+colorTo: red
 sdk: gradio
-sdk_version: 6.31.0
-python_version: '3.13'
+sdk_version: 4.44
 app_file: app.py
 pinned: false
+license: mit
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# 🌡️ AIRA v2 — Arid-climate Intelligent Risk Assistant
+
+**AI-powered voice & vision safety copilot for heat-exposed workers in Qatar.**
+Built for the **QAI × Hugging Face Hackathon**. Aligned with **Qatar National Vision 2030 — Quality of Life**.
+
+## 🏗️ Architecture — Full Qwen Stack
+
+| Layer | Model | Role |
+|-------|-------|------|
+| 👁️ **Vision** | `Qwen/Qwen2-VL-2B-Instruct` | Image analysis of job-site heat-stress indicators |
+| 🧠 **Reasoning** | `Qwen/Qwen3-4B-Instruct` | WBGT calculation, hydration & work-rest planning (Qatar MoPH aligned) |
+| 🎤 **STT** | `openai/whisper-tiny` | Speech-to-text in Arabic / Hindi / Urdu / English |
+| 🔊 **TTS** | `gTTS` | Spoken safety advice read back to worker |
+| 🌤️ **Weather** | Open-Meteo API | Live Qatar temperature & humidity |
+
+## 📁 Repository Structure
+
+```
+AIRA/
+├── app.py              # Gradio Space: Vision + Qwen3 Reasoning + Voice + WBGT Safety
+├── train_qwen_vision.py # LoRA fine-tuning script for Qwen2-VL on heat-stress imagery
+├── requirements.txt    # Python dependencies
+└── README.md           # This file (HF Space metadata + docs)
+```
+
+## 🚀 Deploy to Hugging Face Spaces
+
+### Prerequisites
+> ⚠️ Gradio Spaces require a **HF Pro subscription** on cpu-basic hardware.  
+> Subscribe at https://huggingface.co/pro (your $100 credits can cover this).
+
+### One-command deploy
+
+```bash
+# Install HF CLI
+curl -LsSf https://hf.co/cli/install.sh | bash -s
+
+# Login (you must do this yourself)
+hf auth login
+
+# Create + upload the Space (run once Pro is active)
+hf repos create jonathan4e/aira-v2 --type space --sdk gradio --exist-ok
+hf upload jonathan4e/aira-v2 --type space app.py requirements.txt README.md --commit-message "Deploy AIRA v2"
+hf spaces settings jonathan4e/aira-v2 --hardware t4-small
+hf spaces wait jonathan4e/aira-v2 --timeout 600
+```
+
+**Your Space URL:** `https://huggingface.co/spaces/jonathan4e/aira-v2`
+
+## 🎯 Fine-tune Qwen2-VL on Your Own Dataset
+
+```bash
+# 1. Create/push a heat-stress image dataset to HF
+python create_dataset.py --push jonathan4e/aira-heat-stress
+
+# 2. Fine-tune Qwen2-VL-2B with LoRA (PEFT)
+python train_qwen_vision.py \
+  --dataset jonathan4e/aira-heat-stress \
+  --output jonathan4e/aira-qwen2vl-heat \
+  --epochs 1 --batch 2
+
+# 3. Update app.py VISION_MODEL_ID to your fine-tuned model
+```
+
+## 🧠 WBGT Safety Logic (Qatar MoPH Aligned)
+
+- **WBGT estimate** from temperature + humidity + solar load
+- **Risk bands:** Low (<29°C), Moderate (29–31°C), High (31–33°C), Extreme (>33°C)
+- **Output:** hydration (L/hr), work/rest cycle (min work / min rest shade), immediate actions
+- Multilingual: English, Arabic, Hindi, Urdu
+
+## 📱 Demo Features
+
+- `gr.Examples` pre-loaded for instant hackathon judging
+- Mobile-responsive Gradio UI with RTL support for Arabic/Urdu
+- Voice input → transcription → Qwen3 reasoning → TTS audio output
+- Image upload → Qwen2-VL heat-stress classification
+
+## 🇶🇦 Alignment
+
+Supports **Qatar National Vision 2030 — Quality of Life** by protecting outdoor, logistics, and construction workers from extreme summer heat with free, open-source, mobile-accessible AI.
+
+## 📄 License
+
+MIT

app.py

@@ -0,0 +1,544 @@
+"""
+app.py — AIRA: Arid-climate Intelligent Risk Assistant
+Complete Gradio app for Hugging Face Spaces.
+Qwen stack: Qwen2-VL-2B (vision) + Qwen3-8B-Instruct (reasoning) + Whisper-Tiny (STT) + gTTS (TTS)
+"""
+from __future__ import annotations
+
+import math
+import threading
+import requests
+from datetime import datetime, timezone
+from typing import Optional, Dict, Any, List, Tuple
+
+import gradio as gr
+from PIL import Image, ImageDraw
+
+# ---------------- Optional ML deps (lazy load) ----------------
+try:
+    import torch
+    from transformers import (
+        Qwen2VLForConditionalGeneration,
+        AutoProcessor,
+        AutoModelForCausalLM,
+        AutoTokenizer,
+        pipeline,
+    )
+    _TORCH = True
+except Exception:
+    _TORCH = False
+
+try:
+    from gtts import gTTS
+    _GTTS = True
+except Exception:
+    _GTTS = False
+
+# ---------------- Config ----------------
+QATAR_CITIES: Dict[str, Dict[str, float]] = {
+    "Doha": {"lat": 25.2867, "lon": 51.5343},
+    "Al Wakrah": {"lat": 25.1715, "lon": 51.6047},
+    "Al Khor": {"lat": 25.6819, "lon": 51.5010},
+    "Mesaieed": {"lat": 24.7728, "lon": 51.5045},
+    "Dukhan": {"lat": 25.4219, "lon": 50.7531},
+    "Al Rayyan": {"lat": 25.2333, "lon": 51.4333},
+    "Lusail": {"lat": 25.3628, "lon": 51.5010},
+}
+
+LANGS: Dict[str, Dict[str, Any]] = {
+    "en": {"name": "English", "tts": "en", "rtl": False},
+    "ar": {"name": "\u0627\u0644\u0639\u0631\u0628\u064a\u0629", "tts": "ar", "rtl": True},
+    "hi": {"name": "\u0939\u093f\u0928\u094d\u0926\u093f", "tts": "hi", "rtl": False},
+    "ur": {"name": "\u0627\u0631\u062f\u0648", "tts": "ur", "rtl": False},
+}
+
+VISION_MODEL_ID = "Qwen/Qwen2-VL-2B-Instruct"
+REASON_MODEL_ID = "Qwen/Qwen3-4B-Instruct"   # Qwen3 reasoning engine
+ASR_MODEL_ID = "openai/whisper-tiny"
+
+VISION_PROMPT = (
+    "You are AIRA, a workplace heat-safety assistant for Qatar outdoor workers. "
+    "Inspect the image for heat-stress indicators: sweat, flushed/pale skin, dizziness posture, "
+    "lack of shade, direct sun on workers, missing water, heavy PPE, hot surfaces, unsafe conditions. "
+    "Classify into: Safe_Shaded, Direct_Sun_Exposure, High_Physical_Exertion, Visible_Heat_Distress. "
+    "Be concise, practical, safety-first. Reply in the user's language."
+)
+
+QATAR_SYSTEM_PROMPT = (
+    "You are AIRA, a heat-safety reasoning engine for Qatar outdoor workers, aligned with Qatar "
+    "Ministry of Public Health (MoPH) Wet Bulb Globe Temperature (WBGT) guidelines. "
+    "Given temperature, humidity, work intensity, and vision findings, output: "
+    "(1) WBGT estimate, (2) Heat risk level (Low/Moderate/High/Extreme), "
+    "(3) Hydration plan (Litres per hour), (4) Work-rest cycle (minutes work / minutes rest in shade), "
+    "(5) Immediate safety actions in Arabic, Hindi, Urdu, or English as requested. "
+    "Be concise, actionable, and safety-first."
+)
+
+# Model singletons (lazy, thread-safe)
+_lock = threading.Lock()
+_vm = None
+_vp = None
+_rm = None
+_rt = None
+_asr = None
+
+
+def _load_vision():
+    global _vm, _vp
+    with _lock:
+        if _vm is None and _TORCH:
+            _vm = Qwen2VLForConditionalGeneration.from_pretrained(
+                VISION_MODEL_ID,
+                torch_dtype=torch.bfloat16,
+                device_map="auto",
+                trust_remote_code=True,
+            )
+            _vp = AutoProcessor.from_pretrained(VISION_MODEL_ID, trust_remote_code=True)
+    return _vm, _vp
+
+
+def _load_reason():
+    global _rm, _rt
+    with _lock:
+        if _rm is None and _TORCH:
+            _rm = AutoModelForCausalLM.from_pretrained(
+                REASON_MODEL_ID,
+                torch_dtype=torch.bfloat16,
+                device_map="auto",
+                trust_remote_code=True,
+            )
+            _rt = AutoTokenizer.from_pretrained(REASON_MODEL_ID, trust_remote_code=True)
+    return _rm, _rt
+
+
+def _load_asr():
+    global _asr
+    with _lock:
+        if _asr is None and _TORCH:
+            _asr = pipeline(
+                "automatic-speech-recognition",
+                model=ASR_MODEL_ID,
+                device=0 if torch.cuda.is_available() else -1,
+            )
+    return _asr
+# ---------------- Weather (Open-Meteo, free, no key) ----------------
+def get_qatar_weather(city: str = "Doha") -> Dict[str, Any]:
+    c = QATAR_CITIES.get(city, QATAR_CITIES["Doha"])
+    try:
+        r = requests.get(
+            "https://api.open-meteo.com/v1/forecast",
+            params={
+                "latitude": c["lat"],
+                "longitude": c["lon"],
+                "current": "temperature_2m,relative_humidity_2m,apparent_temperature,wind_speed_10m",
+                "timezone": "Asia/Qatar",
+            },
+            timeout=15,
+        )
+        r.raise_for_status()
+        d = r.json()["current"]
+        return {
+            "city": city,
+            "temp": d.get("temperature_2m"),
+            "feels_like": d.get("apparent_temperature"),
+            "humidity": d.get("relative_humidity_2m"),
+            "wind": d.get("wind_speed_10m"),
+            "ok": True,
+        }
+    except Exception as e:
+        return {"city": city, "temp": 44, "feels_like": 52, "humidity": 45, "wind": 12, "ok": False, "e": str(e)}
+
+
+# ---------------- WBGT / Safety Engine (Qatar MoPH aligned) ----------------
+def _wet_bulb(temp_c: float, rh: float) -> float:
+    """Stull approximation of wet-bulb temperature (°C)."""
+    t = temp_c
+    h = rh
+    tw = (
+        t * math.atan(0.151977 * math.sqrt(h + 8.313659))
+        + math.atan(t + h)
+        - math.atan(h - 1.676331)
+        + 0.00391838 * h ** (3 / 2) * math.atan(0.023101 * h)
+        - 4.686035
+    )
+    return tw
+
+
+def wbgt_estimate(temp_c: float, rh: float, sun_exposed: bool = True) -> float:
+    """Estimate WBGT (°C). If sun-exposed, add solar radiation load (~+2 to +4°C globe)."""
+    tw = _wet_bulb(temp_c, rh)
+    # Simplified WBGT = 0.7*Tw + 0.2*Tg + 0.1*T ; approximate Tg ≈ T + solar load
+    tg = temp_c + (3.0 if sun_exposed else 1.0)
+    wbgt = 0.7 * tw + 0.2 * tg + 0.1 * temp_c
+    return round(wbgt, 1)
+
+
+def wbgt_guidance(wbgt: float, intensity: str = "moderate") -> Dict[str, Any]:
+    """Qatar MoPH-style work/rest & hydration guidance from WBGT."""
+    if wbgt < 29:
+        level, label, color = "low", "Low", "#22c55e"
+        work, rest = 60, 0
+        water = 0.5
+    elif wbgt < 31:
+        level, label, color = "moderate", "Moderate", "#eab308"
+        work, rest = 45, 15
+        water = 0.75
+    elif wbgt < 33:
+        level, label, color = "high", "High", "#f97316"
+        work, rest = 30, 30
+        water = 1.0
+    else:
+        level, label, color = "extreme", "Extreme", "#ef4444"
+        work, rest = 0, 60
+        water = 1.25
+
+    # Adjust for intensity
+    mult = {"light": 0.9, "moderate": 1.0, "heavy": 1.2}.get(intensity, 1.0)
+    water = round(water * mult, 2)
+
+    return {
+        "wbgt": wbgt,
+        "level": level,
+        "label": label,
+        "color": color,
+        "work_min": work,
+        "rest_min": rest,
+        "water_l_hr": water,
+        "intensity": intensity,
+    }
+
+
+def safety_report(temp: float, hum: float, intensity: str, vision_text: str, lang: str) -> str:
+    """Generate localized safety report using Qwen3 reasoning (fallback to template)."""
+    wbgt = wbgt_estimate(temp, hum, sun_exposed="direct sun" in vision_text.lower())
+    guidance = wbgt_guidance(wbgt, intensity)
+
+    rm, rt = _load_reason()
+    if rm is not None and rt is not None:
+        try:
+            user_msg = (
+                f"Temperature: {temp}°C, Humidity: {hum}%, WBGT: {wbgt}°C, "
+                f"Intensity: {intensity}, Vision: {vision_text}. "
+                f"Give a concise heat-safety report in {lang}."
+            )
+            messages = [
+                {"role": "system", "content": QATAR_SYSTEM_PROMPT},
+                {"role": "user", "content": user_msg},
+            ]
+            text = rt.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+            inputs = rt(text=[text], return_tensors="pt").to(rm.device)
+            out = rm.generate(**inputs, max_new_tokens=300, do_sample=False)
+            result = rt.batch_decode(out, skip_special_tokens=True)[0]
+            if "assistant" in result:
+                result = result.split("assistant")[-1].strip()
+            return result or _template_report(guidance, vision_text, lang)
+        except Exception as e:
+            return f"(Qwen3 reasoning unavailable: {e})\n\n" + _template_report(guidance, vision_text, lang)
+    return _template_report(guidance, vision_text, lang)
+
+
+def _template_report(guidance: Dict, vision_text: str, lang: str) -> str:
+    templates = {
+        "en": (
+            f"**WBGT:** {guidance['wbgt']}°C  |  **Risk:** {guidance['label']}  \n"
+            f"**Hydration:** {guidance['water_l_hr']} L/hr  |  **Work:** {guidance['work_min']} min / "
+            f"**Rest:** {guidance['rest_min']} min in shade  \n"
+            f"**Vision findings:** {vision_text}  \n"
+            f"**Action:** {guidance['note'] if 'note' in guidance else 'Monitor worker closely.'}"
+        ),
+        "ar": (
+            f"**WBGT:** {guidance['wbgt']}°C  |  **المخاطر:** {guidance['label']}  \n"
+            f"**الترطيب:** {guidance['water_l_hr']} لتر/ساعة  |  **العمل:** {guidance['work_min']} دقيقة / "
+            f"**الراحة:** {guidance['rest_min']} دقيقة في الظل  \n"
+            f"**ملاحظات الرؤية:** {vision_text}"
+        ),
+        "hi": (
+            f"**WBGT:** {guidance['wbgt']}°C  |  **जोखिम:** {guidance['label']}  \n"
+            f"**हाइड्रेशन:** {guidance['water_l_hr']} L/घंटा  |  **काम:** {guidance['work_min']} मिनट / "
+            f"**आराम:** {guidance['rest_min']} मिनट छांव में  \n"
+            f"**विज़न निष्कर्ष:** {vision_text}"
+        ),
+        "ur": (
+            f"**WBGT:** {guidance['wbgt']}°C  |  **خطرہ:** {guidance['label']}  \n"
+            f"**پانی:** {guidance['water_l_hr']} L/گھنٹہ  |  **کام:** {guidance['work_min']} منٹ / "
+            f"** आरام:** {guidance['rest_min']} منٹ سائے میں  \n"
+            f"**بینائی کے نتائج:** {vision_text}"
+        ),
+    }
+    return templates.get(lang, templates["en"])
+# ---------------- Vision analysis (Qwen2-VL) ----------------
+def analyze_image(image: Image.Image, lang: str = "en") -> str:
+    vm, vp = _load_vision()
+    if vm is None or vp is None:
+        return _vision_fallback(image, lang)
+    try:
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "image", "image": image},
+                    {"type": "text", "text": VISION_PROMPT},
+                ],
+            }
+        ]
+        text = vp.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+        inputs = vp(text=[text], images=[image], padding=True, return_tensors="pt").to(vm.device)
+        out = vm.generate(**inputs, max_new_tokens=256, do_sample=False)
+        result = vp.batch_decode(out, skip_special_tokens=True)[0]
+        if "assistant" in result:
+            result = result.split("assistant")[-1].strip()
+        return result or _vision_fallback(image, lang)
+    except Exception as e:
+        return f"(Vision error: {e})\n\n" + _vision_fallback(image, lang)
+
+
+def _vision_fallback(image: Image.Image, lang: str) -> str:
+    wd, ht = image.size
+    msgs = {
+        "en": f"Vision fallback ({wd}x{ht}). Check sweat, skin, shade, water. If unwell, stop & cool down.",
+        "ar": f"\u0628\u062f\u064a\u0644 \u0627\u0644\u0631\u0624\u064a\u0629 ({wd}\u00d7{ht}). \u062a\u062d\u0642\u0642 \u0645\u0646 \u0627\u0644\u0639\u0631\u0642, \u0627\u0644\u0628\u0634\u0627\u0631\u0629, \u0627\u0644\u0638\u0644, \u0627\u0644\u0645\u0627\u0621.",
+        "hi": f"\u0935\u093f\u091c\u094d\u091e\u093e \u092c\u093e\u0902\u0921 ({wd}\u00d7{ht}). \u092a\u0938\u094d\u0925\u093e\u0928, \u0924\u094d\u092f\u093e\u0916, \u091b\u092a\u093e\u0930, \u092a\u093e\u0928\u093e \u091c\u093e\u0924\u094d\u092e\u093e.",
+        "ur": f"\u0628\u06cc\u0646\u0627\u0626\u06cc \u0645\u062a\u0628\u0627\u062f\u0644 ({wd}\u00d7{ht}). \u067e\u0633\u06cc\u0646\u0627, \u062c\u0644\u062f, \u0633\u0627\u064a\u0647, \u067e\u0627\u0646\u06cc \u062a\u0627\u0645 \u06a9\u06cc\u0646.",
+    }
+    return msgs.get(lang, msgs["en"])
+
+
+# ---------------- Voice: STT + TTS ----------------
+def transcribe(audio_path: Optional[str], lang: str) -> str:
+    if not audio_path:
+        return ""
+    asr = _load_asr()
+    if asr is None:
+        return "(Whisper unavailable — type instead.)"
+    try:
+        res = asr(audio_path, generate_kwargs={"language": lang if lang == "en" else None})
+        return res.get("text", "").strip()
+    except Exception as e:
+        return f"(STT error: {e})"
+
+
+def tts_to_file(text: str, lang: str) -> Optional[str]:
+    if not text or not _GTTS:
+        return None
+    try:
+        path = "/tmp/aira_tts.mp3"
+        gTTS(text=text, lang=LANGS.get(lang, {}).get("tts", "en")).save(path)
+        return path
+    except Exception:
+        return None
+
+# ---------------- UI strings ----------------
+UI: Dict[str, Dict[str, str]] = {
+    "en": {
+        "t": "AIRA v2", "s": "Voice • Vision • Safety Copilot — Qatar Heat Workers",
+        "tag": "Qwen2-VL • Qwen3 Reasoning • Whisper STT • gTTS • HF Spaces",
+        "w": "Live Qatar Weather", "c": "City", "rf": "Refresh",
+        "v": "Vision Scan", "vo": "Voice Copilot", "sf": "Safety Planner",
+        "vp": "Concern? (optional)", "vb": "Analyze Image",
+        "vb2": "Record & Translate", "tts": "Speak Safety Advice",
+        "sn": "Hydration & Work-Rest Plan", "wt": "Weight (kg)",
+        "wi": "Work intensity", "l": "Light", "m": "Moderate", "h": "Heavy",
+        "lg": "Language", "st": "Temp (°C)", "hu": "Humidity (%)",
+        "wbgt": "WBGT & Safety", "f": "Qatar Vision 2030 • Quality of Life • Open Source",
+    },
+    "ar": {
+        "t": "أيرا v2", "s": "مساعد صوت وبصر وأمان — عمال حرارة قطر",
+        "tag": "كيوين2-في إل • كيوين3 reasoning • ويسبر • جي تي تي إس • هجنج فيس سبيسز",
+        "w": "\u0627\u0644\u0637\u0642\u0633", "c": "\u0627\u0644\u0645\u062f\u064a\u0646\u0629", "rf": "\u062a\u062d\u062f\u064a\u062f",
+        "v": "\u062a\u062d\u0644\u064a\u0644 \u0628\u0635\u0631\u064a", "vo": "\u0627\u0644\u0635\u0648\u062a", "sf": "\u0627\u0644\u0623\u0645\u0646",
+        "vp": "\u0645\u0627\u0630\u0627 \u064a\u0642\u0644\u0642\u0643? ", "vb": "\u062a\u062d\u0644\u064a\u0644 \u0627\u0644\u0635\u0648\u0627\u0631",
+        "vb2": "\u062a\u0633\u062c\u064a\u0644 \u0648\u062a\u0631\u062c\u0645\u0629", "tts": "\u0627\u0633\u062a\u0645\u0639 \u0644\u0644\u0646\u0635\u064a\u062d\u0629",
+        "sn": "\u062e\u0637\u0629 \u0627\u0644\u062a\u0631\u0637\u064a\u0628 \u0648\u0627\u0644\u0648\u0642\u0641", "wt": "\u0627\u0644\u0648\u0632\u0646 (\u0643\u063a)",
+        "wi": "\u0634\u062f\u0629 \u0627\u0644\u0639\u0645\u0644", "l": "\u062e\u0641\u064a\u0641", "m": "\u0645\u062a\u0648\u0633\u0637", "h": "\u062b\u0642\u064a\u0644",
+        "lg": "\u0627\u0644\u0644\u063a\u0629", "st": "\u062f\u0631\u062c\u0629 \u0627\u0644\ud83e\udd21\ufe0f", "hu": "\u0646\u0633\u0628\u0629 \u0627\u0644\u0647\u0648\u0627\u0621 (%)",
+        "wbgt": "\u0648\u0628\u062c\u064a\u0634\u064a \u0648\u0627\u0644\u0623\u0645\u0646", "f": "\u0631\u0624\u064a\u0629 \u0642\u0637\u0631 2030 \u2022 \u0645\u062d\u062a\u0648\u0649 \u0627\u0644\u062d\u064a\u0627\u0629 \u2022 \u0645\u0641\u062a\u0648\u062d \u0627\u0644\u0645\u062d\u062f\u0648",
+    },
+    "hi": {
+        "t": "आईरा v2", "s": "आवाज़ • दृष्टि • सुरक्षा सहायक — कतर के गर्मी श्रमिक",
+        "tag": "क्यूइन2-वीएल • क्यूइन3 तर्क • व्हिस्पर • जीटीटीएस • एचएफ स्पेस",
+        "w": "\u092e\u094c\u092e\u093e", "c": "\u0936\u0939\u0930", "rf": "\u0924\u093e\u091a\u093e\u0921\u093c",
+        "v": "\u0935\u093f\u091c\u094d\u091e\u093e \u0938\u0915\u092c\u0942", "vo": "\u0906\u0935\u094d\u092f\u093e", "sf": "\u092f\u0947\u0928\u094d\u0927\u0915\u093e\u0930\u094d\u091c",
+        "vp": "\u091a\u094d\u092f\u093e\u0924\u093f? ", "vb": "\u091b\u093e\u0935\u093f \u092a\u094d\u0930\u0935\u093e\u0928",
+        "vb2": "\u0930\u093f\u0902\u0915\u094d\u0917 \u640 \u0905\u0928\u0941\u0926\u093e\u0921", "tts": "\u0938\u092e\u093e \u0938\u0942\u0928\u093e",
+        "sn": "\u091c\u0932\u094d\u092f\u094b\u092b\u094c\u0928 \u2022 \u0935\u093e\u0930\u094d\u092f\u093e-\u0935\u094d\u0930\u092e\u093f\u0930\u094d\u092e\u093e", "wt": "\u0935\u0928\u0928 (\u0915\u093f\u0902\u0917)",
+        "wi": "\u0915\u093e\u0930\u094d\u0924 \u0924\u093f\u0935\u093e\u091f", "l": "\u0939\u092e\u0932", "m": "\u092e\u0902\u0921\u0935\u094d\u092e\u094d\u092e", "h": "\u092d\u093e\u0930\u0940",
+        "lg": "\u092d\u093e\u0907\u091c", "st": "\u0926\u094d\u0935\u093e\u091f (\u092e\u094d\u092e\u093e\u0902)", "hu": "\u0928\u094d\u092f\u093e\u0930\u094d\u0925\u093e\u092f (%)",
+        "wbgt": "डब्ल्यूबीजीटी और सुरक्षा", "f": "\u0915\u093e\u0924\u0930 \u0935\u093f\u0936\u094d\u0935\u0935\u093e\u0928 2030 \u2022 \u091c\u0940\u092e\u093e\u0928 \u0917\u0941\u0928\u093f\u092e \u2022 \u0914\u092a\u0923 \u0938\u094c\u092e\u0938",
+    },
+    "ur": {
+        "t": "ایرا v2", "s": "آواز • بینائی • حفاظتی معاون — گرمی کے مزدور",
+        "tag": "کوئیں2-وی ایل • کوئیں3 استدلال • ويسپر • جی ٹی ٹی ایس • ایچ ایف سلیٹس",
+        "w": "\u0645\u0648\u0633\u0645", "c": "\u0634\u0647\u0631", "rf": "\u062a\u0627\u0632\u0647 \u0643\u06cc\u064a\u0627",
+        "v": "\u0628\u06cc\u0646\u0627\u0626\u06cc", "vo": "\u0622\u0627\u0632", "sf": "\u067e\u0644\u0627\u0646 \u0627\u0645\u0646",
+        "vp": "\u06a9\u06cc\u0627 \u0646\u06cc\u06ba\u0634? ", "vb": "\u062a\u0633\u0648\u06cc\u0631 \u0635\u0648\u06cc\u0631",
+        "vb2": "\u0631\u06cc\u06a9\u0631\u0627\u062f \u0648 \u062a\u0631\u062c\u0645\u0629", "tts": "\u0645\u0634\u0648\u0631\u0647 \u0633\u0646\u06cc\u0646",
+        "sn": "\u067e\u0627\u0646\u06cc \u0648 \u0648\u0642\u0641\u0647", "wt": "\u0648\u0632\u0646 (\u06a9\u06cc\u0644\u0648\u063a\u0631\u0627\u0645)",
+        "wi": "\u06a9\u0627\u0645 \u0643\u06cc \u0634\u062f\u062a", "l": "\u0647\u0644\u0643\u0627", "m": "\u0645\u062a\u0648\u0633\u0648\u062a", "h": "\u0628\u062d\u0627\u0631\u06cc",
+        "lg": "\u0632\u0628\u0627\u0646", "st": "درجہ حرارت (\u0628\u0648)", "hu": "نمی (\u0641\u06ccصد)",
+        "wbgt": "وی بی جی ٹی اور حفاظت", "f": "\u0642\u0637\u0631 \u0648\u06cc\u0632\u06cc\u0646 2030 \u2022 \u0632\u06cc\u0646\u06af\u06cc \u0643\u0627 \u0645\u06cc\u0627\u0631 \u2022 \u0627\u0648\u067e\u0646 \u0633\u0648\u0631\u0633",
+    },
+}
+
+
+def tr(lang: str, k: str) -> str:
+    return UI.get(lang, UI["en"]).get(k, UI["en"][k])
+
+# ---------------- CSS ----------------
+CSS = """
+:root{--deep:#78350f;--desert:#d97706;--maroon:#8d2f3a;}
+.gradio-container{font-family:system-ui,sans-serif;background:linear-gradient(135deg,#fdf3e3 0%,#e8d4b3 60%,#d97706 100%);min-height:100vh}
+body[dir=rtl],.gradio-container[dir=rtl]{direction:rtl}
+.aira-header{text-align:center;padding:2rem 1rem 1rem;border-bottom:3px solid var(--maroon);margin-bottom:1rem}
+.aira-title{font-size:2.5rem;font-weight:800;color:var(--deep);margin:0}
+.aira-subtitle{font-size:1.1rem;color:#374151;margin:.25rem 0 .5rem}
+.aira-tagline{font-size:.9rem;color:var(--desert);font-weight:600;text-transform:uppercase;letter-spacing:1px}
+.weather-card{background:rgba(255,255,255,.85);border-radius:16px;padding:1.25rem;box-shadow:0 10px 25px -5px rgba(0,0,0,.1);margin-bottom:1rem;border-left:5px solid var(--maroon)}
+.weather-temp{font-size:3rem;font-weight:900;color:var(--deep)}
+.risk-badge{display:inline-block;padding:.35rem 1rem;border-radius:999px;font-weight:800;font-size:.95rem;color:#fff}
+.plan-card{background:rgba(255,255,255,.85);border-radius:16px;padding:1.25rem;box-shadow:0 10px 25px -5px rgba(0,0,0,.1);margin-bottom:1rem}
+footer{text-align:center;padding:1.5rem;color:#6b7280;font-size:.85rem;border-top:1px solid rgba(255,255,255,.3);margin-top:2rem}
+@media(max-width:640px){.aira-title{font-size:1.75rem}.weather-temp{font-size:2.25rem}}
+# ---------------- Gradio App ----------------
+def build_app() -> gr.Blocks:
+    with gr.Blocks(title="AIRA v2", css=CSS, theme=gr.themes.Soft(primary_hue="orange")) as demo:
+        # State
+        lang = gr.State("en")
+        city = gr.State("Doha")
+        ws = gr.State({})
+        guidance = gr.State({})
+
+        # Header
+        gr.HTML(
+            f'<div class="aira-header">'
+            f'<div class="aira-title">\U0001f321\ufe0f {tr("en","t")}</div>'
+            f'<div class="aira-subtitle">{tr("en","s")}</div>'
+            f'<div class="aira-tagline">{tr("en","tag")}</div>'
+            f'</div>'
+        )
+
+        # Weather row
+        with gr.Row(equal_height=True):
+            cd = gr.Dropdown(choices=list(QATAR_CITIES), value="Doha", label=tr("en","c"), scale=1)
+            rb = gr.Button(tr("en","rf"), variant="primary", scale=0)
+            wd = gr.HTML(value=_weather_html("Doha", get_qatar_weather("Doha")))
+
+        # Tabs
+        with gr.Tabs():
+            # --- Vision Tab ---
+            with gr.TabItem(tr("en","v"), id="vision"):
+                with gr.Row():
+                    with gr.Column(scale=1):
+                        vi = gr.Image(type="pil", label="Worker / job-site photo")
+                        vp2 = gr.Textbox(label=tr("en","vp"), placeholder="e.g. sweating, no shade...", lines=2)
+                        vb = gr.Button(tr("en","vb"), variant="primary")
+                    with gr.Column(scale=1):
+                        vo = gr.Markdown(label="Aira Vision Assessment")
+
+            # --- Voice Tab ---
+            with gr.TabItem(tr("en","vo"), id="voice"):
+                with gr.Row():
+                    with gr.Column(scale=1):
+                        va = gr.Audio(sources=["microphone"], type="filepath", label="Speak AR/HI/UR/EN")
+                        vb2 = gr.Button(tr("en","vb2"), variant="primary")
+                        vt = gr.Textbox(label="Transcribed", lines=3, interactive=False)
+                    with gr.Column(scale=1):
+                        tb = gr.Button(tr("en","tts"), variant="secondary")
+                        ta = gr.Audio(label="Safety audio (auto-play)", autoplay=True)
+
+            # --- Safety Tab ---
+            with gr.TabItem(tr("en","sf"), id="safety"):
+                with gr.Row():
+                    with gr.Column(scale=1):
+                        st_in = gr.Number(label=tr("en","st"), value=40, minimum=0, maximum=60, step=1)
+                        hu_in = gr.Number(label=tr("en","hu"), value=50, minimum=0, maximum=100, step=1)
+                        wi_in = gr.Radio([tr("en","l"), tr("en","m"), tr("en","h")], value=tr("en","m"), label=tr("en","wi"))
+                        sl_in = gr.Dropdown([f"{x['name']} ({k})" for k, x in LANGS.items()], value="English (en)", label=tr("en","lg"))
+                    with gr.Column(scale=1):
+                        rh_out = gr.HTML(label=tr("en","wbgt"))
+                        ph_out = gr.HTML(label=tr("en","sn"))
+
+        gr.HTML(f"<footer>{tr('en','f')}</footer>")
+
+        # ---------------- Helpers ----------------
+        def _weather_html(ct: str, w: Dict) -> str:
+            if not w.get("ok"):
+                return f'<div class="weather-card"><div class="weather-temp">{w.get("temp","—")}\u00b0C</div><div style="color:#6b7280">{ct} — weather unavailable</div></div>'
+            r = wbgt_guidance(wbgt_estimate(w.get("temp"), w.get("humidity"), True), "moderate")
+            return (
+                f'<div class="weather-card">'
+                f'<div style="display:flex;justify-content:space-between;align-items:start;">'
+                f'<div><div class="weather-temp">{w.get("temp")}\u00b0C</div>'
+                f'<div style="font-size:.9rem;color:#4b5563">Feels {w.get("feels_like")}\u00b0C • Hum {w.get("humidity")}% • Wind {w.get("wind")} km/h • {ct}</div></div>'
+                f'<div class="risk-badge" style="background:{r["color"]}">{r["label"]} risk</div>'
+                f'</div></div>'
+            )
+
+        def _safety_html(temp, hum, intensity, lang_code, vision_text=""):
+            wbgt = wbgt_estimate(temp, hum, sun_exposed="direct sun" in vision_text.lower())
+            g = wbgt_guidance(wbgt, intensity)
+            rep = safety_report(temp, hum, intensity, vision_text, lang_code)
+            html = (
+                f'<div class="plan-card" style="border-left:5px solid {g["color"]}">'
+                f'<div style="display:flex;justify-content:space-between;margin-bottom:.5rem;">'
+                f'<strong>WBGT:</strong> {g["wbgt"]}°C <strong>Risk:</strong> <span class="risk-badge" style="background:{g["color"]}">{g["label"]}</span>'
+                f'</div>'
+                f'<div style="margin-top:.5rem;line-height:1.6;">{rep}</div>'
+                f'</div>'
+            )
+            return html, g
+
+        # ---------------- Event handlers ----------------
+        def oref(c):
+            w = get_qatar_weather(c)
+            ws.value = w
+            return _weather_html(c, w)
+
+        def olg(ch):
+            lang.value = ch.split("(")[-1].strip(")") if "(" in ch else "en"
+
+        def ov(img, pr, l_code, demo_flag):
+            if demo_flag or img is None:
+                return _vision_fallback(Image.new("RGB", (400, 300), (200, 200, 200)), l_code)
+            return analyze_image(img, l_code)
+
+        def ovc(audio, l_code):
+            tx = transcribe(audio, l_code)
+            w = ws.value or get_qatar_weather(city.value or "Doha")
+            g = guidance.value or wbgt_guidance(wbgt_estimate(w.get("temp"), w.get("humidity"), True), "moderate")
+            rep = safety_report(w.get("temp"), w.get("humidity"), g["intensity"], tx or "no voice input", l_code)
+            if tx:
+                rep = f'You said: "{tx}"\n\n{rep}'
+            tts_path = tts_to_file(rep, l_code)
+            return rep, tts_path
+
+        def osafety(temp, hum, wi_label, lang_choice):
+            l_code = lang_choice.split("(")[-1].strip(")") if "(" in lang_choice else "en"
+            lang.value = l_code
+            intensity = {"light": "light", "moderate": "moderate", "heavy": "heavy"}.get(
+                {tr("en","l"):"light", tr("en","m"):"moderate", tr("en","h"):"heavy"}.get(wi_label, "moderate"), "moderate"
+            )
+            html, g = _safety_html(temp, hum, intensity, l_code)
+            guidance.value = g
+            return html
+
+        def ovoice_tts(txt, l_code):
+            return tts_to_file(txt, l_code)
+
+        # Wiring
+        rb.click(oref, [cd], [wd])
+        cd.change(oref, [cd], [wd])
+        sl_in.change(olg, [sl_in], [])
+        vb.click(ov, [vi, vp2, lang, gr.Checkbox(value=False, visible=False)], [vo])
+        vb2.click(ovc, [va, lang], [vt, ta])
+        tb.click(ovoice_tts, [vt, lang], [ta])
+        st_in.change(osafety, [st_in, hu_in, wi_in, sl_in], [rh_out])
+        hu_in.change(osafety, [st_in, hu_in, wi_in, sl_in], [rh_out])
+        wi_in.change(osafety, [st_in, hu_in, wi_in, sl_in], [rh_out])
+        sl_in.change(osafety, [st_in, hu_in, wi_in, sl_in], [rh_out])
+        demo.load(osafety, [st_in, hu_in, wi_in, sl_in], [rh_out])
+
+    return demo
+
+
+if __name__ == "__main__":
+    build_app().launch(server_name="0.0.0.0", server_port=7860)

requirements.txt

@@ -0,0 +1,12 @@
+gradio>=4.44,<6
+transformers>=4.46
+torch>=2.3
+accelerate>=0.33
+qwen-vl-utils>=0.0.8
+datasets>=2.19
+peft>=0.11
+gtts>=2.3
+Pillow>=10
+requests>=2.31
+numpy
+huggingface_hub>=0.24