Spaces:

Armacam
/

agente-h-brido-nvidia-edge-ai-orchestrator

Running

App Files Files Community

Armacam commited on Jul 2

Commit

fc3d107

verified ·

1 Parent(s): 0153f13

Deploy from HuggingChat

Browse files

Files changed (2) hide show

README.md +9 -5
index.html +562 -19

README.md CHANGED Viewed

@@ -1,10 +1,14 @@
 ---
-title: Agente H Brido Nvidia Edge Ai Orchestrator
-emoji: 📉
-colorFrom: yellow
-colorTo: green
 sdk: static
 pinned: false
 ---
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

 ---
+title: "Agente Híbrido NVIDIA - Edge AI Orchestrator"
+emoji: 🤗
+colorFrom: gray
+colorTo: pink
 sdk: static
 pinned: false
+tags:
+  - huggingchat
 ---
+# Agente Híbrido NVIDIA - Edge AI Orchestrator
+Built with [HuggingChat](https://huggingface.co/chat).

index.html CHANGED Viewed

@@ -1,19 +1,562 @@
-<!doctype html>
-<html>
-	<head>
-		<meta charset="utf-8" />
-		<meta name="viewport" content="width=device-width" />
-		<title>My static Space</title>
-		<link rel="stylesheet" href="style.css" />
-	</head>
-	<body>
-		<div class="card">
-			<h1>Welcome to your static Space!</h1>
-			<p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
-			<p>
-				Also don't forget to check the
-				<a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
-			</p>
-		</div>
-	</body>
-</html>

+<!DOCTYPE html>
+<html lang="es">
+<head>
+<meta charset="UTF-8">
+<meta name="viewport" content="width=device-width, initial-scale=1.0">
+<title>Agente Híbrido NVIDIA - Edge AI Orchestrator</title>
+<style>
+  :root {
+    --bg: #0b0f19;
+    --card: #111827;
+    --border: #1f2937;
+    --accent: #76b900;
+    --accent2: #00d4ff;
+    --text: #e5e7eb;
+    --muted: #9ca3af;
+    --danger: #f87171;
+  }
+  * { box-sizing: border-box; }
+  body {
+    margin: 0; font-family: system-ui, -apple-system, 'Segoe UI', Roboto, sans-serif;
+    background: var(--bg); color: var(--text); line-height: 1.6;
+  }
+  header {
+    background: linear-gradient(90deg, #0000%, #0b1d0b 100%);
+    border-bottom: 1px solid var(--border);
+    padding: 2rem 1.5rem; text-align: center;
+  }
+  header h1 { margin: 0; font-size: 1.8rem; color: var(--accent); letter-spacing: -0.5px; }
+  header p { margin: .5rem 0 0; color: var(--muted); font-size: .95rem; }
+  .container { max-width: 1000px; margin: 0 auto; padding: 1.5rem; }
+  .tabs { display: flex; gap: .5rem; flex-wrap: wrap; margin-bottom: 1.5rem; }
+  .tab {
+    background: var(--card); border: 1px solid var(--border);
+    color: var(--muted); padding: .65rem 1.1rem; border-radius: 8px;
+    cursor: pointer; font-size: .92rem; transition: .2s;
+  }
+  .tab:hover { color: var(--text); border-color: var(--accent); }
+  .tab.active { color: #fff; border-color: var(--accent); background: #162818; }
+  .panel { display: none; animation: fade .3s ease; }
+  .panel.active { display: block; }
+  @keyframes fade { from { opacity: 0; transform: translateY(6px); } to { opacity: 1; } }
+  .card {
+    background: var(--card); border: 1px solid var(--border);
+    border-radius: 12px; padding: 1.5rem; margin-bottom: 1.5rem;
+  }
+  .card h2 { margin-top: 0; font-size: 1.15rem; color: var(--accent2); }
+  .tag {
+    display: inline-block; background: #1e293b; color: var(--accent2);
+    padding: .2rem .6rem; border-radius: 99px; font-size: .78rem;
+    margin: .15rem .2rem .15rem 0; border: 1px solid #334155;
+  }
+  pre {
+    background: #0e1525; border: 1px solid #1f2937; border-radius: 8px;
+    padding: 1rem; overflow-x: auto; font-size: .85rem; color: #d1d5db;
+  }
+  .highlight { color: var(--accent); }
+  .note {
+    border-left: 3px solid var(--accent); background: #0f1d0f;
+    padding: .8rem 1rem; border-radius: 0 8px 8px 0; color: #bbf7d0; font-size: .9rem;
+  }
+  .grid-2 { display: grid; grid-template-columns: 1fr 1fr; gap: 1.5rem; }
+  @media (max-width: 720px) { .grid-2 { grid-template-columns: 1fr; } }
+  button.copy {
+    float: right; background: #1f2937; border: 1px solid #334155; color: var(--text);
+    padding: .3rem .6rem; border-radius: 6px; font-size: .75rem; cursor: pointer;
+  }
+  button.copy:hover { background: #334155; }
+  svg.diagram { width: 100%; height: auto; }
+</style>
+</head>
+<body>
+<header>
+  <h1>🧠 Agente Híbrido NVIDIA Nemotron</h1>
+  <p>Edge AI Orchestrator · Razonamiento Multimodal · Multiplataforma ARM/Android · Automatización Agéntica</p>
+</header>
+<div class="container">
+  <div class="tabs" id="tabs">
+    <div class="tab active" onclick="show('vision')">Visión General</div>
+    <div class="tab" onclick="show('arq')">Arquitectura</div>
+    <div class="tab" onclick="show('core')">Core Python</div>
+    <div class="tab" onclick="show('android')">Android ARM</div>
+    <div class="tab" onclick="show('auto')">Auto-Cuentas/APIs</div>
+    <div class="tab" onclick="show('deploy')">Deploy</div>
+  </div>
+  <!-- VISIÓN -->
+  <div id="vision" class="panel active">
+    <div class="card">
+      <h2>🎯 Objetivo</h2>
+      <p>
+        Construir un <strong>agente cognitivo híbrido</strong> que use <strong>NVIDIA Nemotron 3</strong> como núcleo de razonamiento,
+        con capacidad para operar tanto en <em>cloud</em> (infraestructura GPU H100/B200) como en <em>edge</em>
+ (Jetson Thor, DGX Spark, Android ARM con NPU/GPU vía Arm NN / MLLM / Cactus).
+        El agente automatiza:
+ </p>
+      <ul>
+        <li><strong>Búsqueda y descubrimiento de APIs</strong> vía Hugging Face Discover (`hf discover`).</li>
+        <li><strong>Lectura inteligente de documentos</strong> (PDFs, manuales, términos) con OCR + reasoning multimodal.</li>
+        <li><strong>Registro y creación de cuentas</strong> automatizado mediante control de navegador + LLM.</li>
+        <li><strong>Generación de código multiplataforma</strong> (Python, Kotlin, C++, Rust) acelerado con NumPy-SVE en ARM.</li>
+        <li><strong>Traducción y razonamiento multilingüe</strong> (EN/ES/DE/FR/IT/JP/中文/한국어).</li>
+      </ul>
+      <div class="note">💡 <strong>Tip de modelo:</strong> Nemotron-3-Nano-Omni-30B-A3B en NVFP4 cabe en Jetson Thor/RTX 5090 y permite video+audio+texto+razonamiento en un solo modelo.</div>
+    </div>
+    <div class="grid-2">
+      <div class="card">
+        <h2>🔧 Stack Tecnológico</h2>
+        <div>
+          <span class="tag">NVIDIA Nemotron 3</span>
+          <span class="tag">Hugging Face Hub</span>
+          <span class="tag">HF CLI & Skills</span>
+          <span class="tag">vLLM / TensorRT-LLM</span>
+          <span class="tag">Playwright</span>
+          <span class="tag">NumPy + SVE</span>
+          <span class="tag">MLLM / Cactus</span>
+          <span class="tag">Arm NN (Android)</span>
+          <span class="tag">KleidiAI</span>
+          <span class="tag">Kotlin Multiplatform</span>
+ <span class="tag">ExecuTorch</span>
+          <span class="tag">QNN / Hexagon NPU</span>
+        </div>
+      </div>
+      <div class="card">
+        <h2>📐 Patrón de Despliegue</h2>
+        <ul>
+          <li><strong>Cloud:</strong> Nemotron-3-Super-120B o Ultra-550B en vLLM con reasoning on/off.</li>
+          <li><strong>Edge:</strong> Nemotron-3-Nano-Omni-30B en NVFP4 via TensorRT-LLM o llama.cpp.</li>
+          <li><strong>Android:</strong> Modelo cuantizado INT4/W4A16 ejecutado con MLLM + servidor in-app Golang.</li>
+        </ul>
+      </div>
+    </div>
+  </div>
+  <!-- ARQUITECTURA -->
+  <div id="arq" class="panel">
+    <div class="card">
+      <h2>🏗 Diagrama de Arquitectura</h2>
+      <svg class="diagram" viewBox="0 0 800 520" xmlns="http://www.w3.org/2000/svg">
+        <defs>
+          <linearGradient id="g1" x1="0" y1="0" x2="1" y2="1"><stop offset="0%" stop-color="#111827"/><stop offset="100%" stop-color="#0b1222"/></linearGradient>
+        </defs>
+        <rect x="10" y="10" width="780" height="500" rx="14" fill="url(#g1)" stroke="#1f2937" stroke-width="1"/>
+        <!-- Nemotron Core -->
+        <rect x="250" y="40" width="300" height="70" rx="10" fill="#162818" stroke="#76b900" stroke-width="2"/>
+        <text x="400" y="65" fill="#76b900" font-size="13" font-weight="bold" text-anchor="middle">NVIDIA Nemotron 3 Core</text>
+        <text x="400" y="85" fill="#9ca3af" font-size="11" text-anchor="middle">Reasoning · Tool Calling · Multimodal</text>
+        <text x="400" y="100" fill="#9ca3af" font-size="11" text-anchor="middle">NVFP4 / BF16 / FP8</text>
+        <!-- Hugging Face Skills -->
+        <rect x="40" y="160" width="220" height="90" rx="10" fill="#0e1525" stroke="#00d4ff" stroke-width="1"/>
+        <text x="150" y="185" fill="#00d4ff" font-size="12" font-weight="bold" text-anchor="middle">HF CLI + Skills</text>
+        <text x="150" y="205" fill="#9ca3af" font-size="10" text-anchor="middle">hf discover · hf skills add</text>
+        <text x="150" y="220" fill="#9ca3af" font-size="10" text-anchor="middle">API search · MCP Servers</text>
+        <text x="150" y="235" fill="#9ca3af" font-size="10" text-anchor="middle">Jobs · Datasets · Models</text>
+        <!-- Automatización -->
+        <rect x="290" y="160" width="220" height="90" rx="10" fill="#0e1525" stroke="#f87171" stroke-width="1"/>
+        <text x="400" y="185" fill="#f87171" font-size="12" font-weight="bold" text-anchor="middle">Agente de Automatización</text>
+        <text x="400" y="205" fill="#9ca3af" font-size="10" text-anchor="middle">Playwright + LLM</text>
+        <text x="400" y="220" fill="#9ca3af" font-size="10" text-anchor="middle">Doc Parser · OCR · RAG</text>
+        <text x="400" y="235" fill="#9ca3af" font-size="10" text-anchor="middle">Creación de Cuentas</text>
+        <!-- Code Gen -->
+        <rect x="540" y="160" width="220" height="90" rx="10" fill="#0e1525" stroke="#fbbf24" stroke-width="1"/>
+        <text x="650" y="185" fill="#fbbf24" font-size="12" font-weight="bold" text-anchor="middle">Generación de Código</text>
+        <text x="650" y="205" fill="#9ca3af" font-size="10" text-anchor="middle">Multi-idioma · Multiplataforma</text>
+        <text x="650" y="220" fill="#9ca3af" font-size="10" text-anchor="middle">Python · Kotlin · C++ · Rust</text>
+        <text x="650" y="235" fill="#9ca3af" font-size="10" text-anchor="middle">NumPy acelerado (SVE/AVX)</text>
+        <!-- Connectors -->
+ <line x1="150" y1="160" x2="320" y2="110" stroke="#374151" stroke-width="1"/>
+        <line x1="400" y1="160" x2="400" y2="110" stroke="#374151" stroke-width="1"/>
+        <line x1="650" y1="160" x2="480" y2="110" stroke="#374151" stroke-width="1"/>
+        <!-- Edge Layer -->
+        <rect x="40" y="290" width="350" height="90" rx="10" fill="#1a1025" stroke="#a78bfa" stroke-width="1"/>
+        <text x="215" y="315" fill="#a78bfa" font-size="12" font-weight="bold" text-anchor="middle">Edge / On-Device Runtime</text>
+        <text x="215" y="335" fill="#9ca3af" font-size="10" text-anchor="middle">MLLM · Cactus · Arm NN · ExecuTorch</text>
+        <text x="215" y="350" fill="#9ca3af" font-size="10" text-anchor="middle">Android (Kotlin) · Linux ARM · Jetson</text>
+        <text x="215" y="365" fill="#9ca3af" font-size="10" text-anchor="middle">INT4/W4A16 · NPU / GPU / CPU</text>
+        <!-- NumPy Accel -->
+        <rect x="410" y="290" width="350" height="90" rx="10" fill="#1a1025" stroke="#f472b6" stroke-width="1"/>
+        <text x="585" y="315" fill="#f472b6" font-size="12" font-weight="bold" text-anchor="middle">Computación Acelerada</text>
+        <text x="585" y="335" fill="#9ca3af" font-size="10" text-anchor="middle">NumPy + OpenBLAS / MKL / SVE</text>
+        <text x="585" y="350" fill="#9ca3af" font-size="10" text-anchor="middle">CuPy / Numba · CUDA / ROCm</text>
+        <text x="585" y="365" fill="#9ca3af" font-size="10" text-anchor="middle">TensorRT-LLM · vLLM · llama.cpp</text>
+        <line x1="400" y1="250" x2="215" y2="290" stroke="#374151" stroke-width="1"/>
+        <line x1="400" y1="250" x2="585" y2="290" stroke="#374151" stroke-width="1"/>
+ <!-- Data Ingest -->
+        <rect x="40" y="420" width="720" height="60" rx="10" fill="#0f172a" stroke="#334155" stroke-width="1"/>
+        <text x="400" y="445" fill="#94a3b8" font-size="12" font-weight="bold" text-anchor="middle">Entradas: Texto · Imagen · Audio · Video · PDF · Web · APIs</text>
+        <text x="400" y="465" fill="#64748b" font-size="10" text-anchor="middle">OCR (Nano Omni) · ASR (Parakeet) · Traducción · Embeddings</text>
+ <line x1="215" y1="380" x2="300" y2="420" stroke="#374151" stroke-width="1"/>
+        <line x1="585" y1="380" x2="500" y2="420" stroke="#374151" stroke-width="1"/>
+      </svg>
+    </div>
+  </div>
+  <!-- CORE PYTHON -->
+  <div id="core" class="panel">
+    <div class="card">
+      <h2>🐍 Core Python: Nemotron + HF Skills + NumPy SVE</h2>
+      <button class="copy" onclick="copy(this)">Copiar</button>
+      <pre><code><span class="highlight"># orchestrator.py - Motor híbrido del agente</span>
+import os
+from huggingface_hub import hf_hub_download, HfApi
+from transformers import AutoTokenizer, AutoModelForCausalLM
+import numpy as np
+import subprocess
+import json
+<span class="highlight"># 1. CONFIGURACIÓN HÍBRIDA</span>
+class HybridConfig:
+    CLOUD_MODEL = "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"
+    EDGE_MODEL  = "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4"
+    HF_TOKEN = os.getenv("HF_TOKEN")
+    ENABLE_SVE  = True  # Activar kernels SVE en ARM
+<span class="highlight"># 2. CLI WRAPPER PARA SKILLS Y DISCOVER</span>
+class HFAgentCLI:
+    @staticmethod
+    def search_skill(query: str, kind="skill"):
+        <span class="highlight">"""Descubre APIs, skills o MCP servers automáticamente."""</span>
+        cmd = ["hf", "discover", "search", query, "--kind", kind, "--json"]
+        out = subprocess.run(cmd, capture_output=True, text=True)
+        return json.loads(out.stdout) if out.returncode == 0 else {}
+    @staticmethod
+    def add_skill(scope="--global"):
+        subprocess.run(["hf", "skills", "add", scope])
+    @staticmethod
+    def download_model(repo_id: str, filename: str = None):
+        api = HfApi(token=HybridConfig.HF_TOKEN)
+        if filename:
+            return hf_hub_download(repo_id, filename, token=HybridConfig.HF_TOKEN)
+        return repo_id
+<span class="highlight"># 3. REASONING ENGINE</span>
+class NemotronEngine:
+    def __init__(self, model_name: str, device="cuda"):
+        self.tok = AutoTokenizer.from_pretrained(
+            model_name, token=HybridConfig.HF_TOKEN, trust_remote_code=True
+        )
+        self.model = AutoModelForCausalLM.from_pretrained(
+            model_name,
+            token=HybridConfig.HF_TOKEN,
+            device_map="auto",
+            torch_dtype="auto",
+            trust_remote_code=True
+        )
+        self.device = device
+    def reason(self, prompt: str, enable_thinking=True) -&gt; dict:
+        <span class="highlight">"""Genera traza de razonamiento + respuesta final."""</span>
+        messages = [
+            {"role": "system", "content": "enable_thinking=True" if enable_thinking else ""},
+            {"role": "user", "content": prompt}
+        ]
+        inputs = self.tok.apply_chat_template(
+            messages, tokenize=True, return_tensors="pt", add_generation_prompt=True
+        ).to(self.device)
+        outputs = self.model.generate(inputs, max_new_tokens=2048, do_sample=False)
+        text = self.tok.decode(outputs[0], skip_special_tokens=False)
+        <span class="highlight"># Parsing de reasoning trace vs final answer</span>
+        return {"raw": text, "thinking": self._extract_thinking(text), "answer": self._extract_answer(text)}
+    def _extract_thinking(self, text): return text.split("&lt;think&gt;")[-1].split("&lt;/think&gt;")[0] if "&lt;think&gt;" in text else ""
+ def _extract_answer(self, text): return text.split("&lt;/think&gt;")[-1] if "&lt;/think&gt;" in text else text
+<span class="highlight"># 4. NUMPY ACELERADO PARA ARM (SVE)</span>
+class AcceleratedCompute:
+    def __init__(self):
+        self.backend = "numpy+openblas"
+        if HybridConfig.ENABLE_SVE:
+            <span class="highlight"># En compilaciones con SVE, NumPy usa kernels SVE para HPC ARM</span>
+            try:
+                import scipy; self.backend += "+scipy"
+            except Exception: pass
+    def fast_matmul(self, a: np.ndarray, b: np.ndarray) -&gt; np.ndarray:
+        <span class="highlight">"""GEMM acelerado; en ARM con SVE puede alcanzar hasta 1300x vs scalar."""</span>
+        return np.dot(a, b)
+<span class="highlight"># 5. API DE AGENTE COMPLETA</span>
+if __name__ == "__main__":
+    <span class="highlight"># Instalar skills primero: hf skills add --global</span>
+    agent = NemotronEngine(HybridConfig.CLOUD_MODEL)
+    compute = AcceleratedCompute()
+    <span class="highlight"># Ejemplo: buscar APIs y razonar sobre ellas</span>
+    skills = HFAgentCLI.search_skill("API para generación de imágenes")
+    print(skills)
+    result = agent.reason(
+        "Analiza este JSON de APIs y dime cuál es mejor para generar imágenes en edge: " + json.dumps(skills),
+        enable_thinking=True
+    )
+    print(result["thinking"])
+    print(result["answer"])
+</code></pre>
+    </div>
+    <div class="note">
+      💡 <strong>Punto clave:</strong> `hf discover search` permite que el agente encuentre automáticamente MCP servers y skills sin intervención humana.
+    </div>
+  </div>
+  <!-- ANDROID ARM -->
+  <div id="android" class="panel">
+    <div class="card">
+      <h2>📱 Android ARM - Edge Inference (MLLM / Cactus / ExecuTorch)</h2>
+      <button class="copy" onclick="copy(this)">Copiar</button>
+      <pre><code><span class="highlight">// build.gradle.kts - Dependencias para edge inference</span>
+dependencies {
+    <span class="highlight">// Opción A: MLLM (Motor de inferencia multi-backend con servidor in-app)</span>
+    implementation(files("libs/mllm_server.aar"))
+    <span class="highlight">// Opción B: Cactus Compute (Kotlin Multiplatform + NPU)</span>
+    implementation("com.cactuscompute:geniex-android:0.2.0")
+    <span class="highlight">// Opción C: ExecuTorch (PyTorch Edge)</span>
+    implementation("org.pytorch:executorch-android:0.5.0")
+}
+<span class="highlight">// HybridAgent.kt - Agente que corre en Android/ARM</span>
+package com.hybridagent.edge
+import android.content.Context
+import kotlinx.coroutines.*
+class HybridEdgeAgent(private val ctx: Context) {
+    private val dispatcher = Dispatchers.Default    <span class="highlight">// Carga modelo cuantizado W4A16 o INT4 para edge</span>
+    suspend fun initModel(modelPath: String) = withContext(dispatcher) {
+        <span class="highlight">// Ejemplo con MLLM: servidor in-app Golang vía FFI</span>
+        MLLMNative.loadModel(modelPath, backend = "qnn") <span class="highlight">// qnn, opencl, cpu</span>
+    }
+    <span class="highlight">// Razonamiento on-device con Nemotron-3-Nano (cuantizado)</span>
+    suspend fun reason(prompt: String): String = withContext(dispatcher) {
+        val system = "enable_thinking=True"
+        val json = """{"messages":[{"role":"system","content":"$system"},{"role":"user","content":"$prompt"}]}"""
+        MLLMNative.chat(json)
+    }
+    <span class="highlight">// OCR / Análisis de documentos (Nano Omni multimodal)</span>
+    suspend fun analyzeDocument(imageBytes: ByteArray): String = withContext(dispatcher) {
+        MLLMNative.processImage(imageBytes, task = "ocr+reasoning")
+    }
+    <span class="highlight">// Traducción acelerada por NPU</span>
+    suspend fun translate(text: String, targetLang: String): String = withContext(dispatcher) {
+        <span class="highlight">// El mismo modelo Nemotron soporta múltiples idiomas nativamente</span>
+        reason("Traduce al $targetLang: $text")
+    }
+    companion object {
+        <span class="highlight">// Singleton para la app</span>
+        @Volatile private var instance: HybridEdgeAgent? = null
+        fun getInstance(ctx: Context): HybridEdgeAgent =
+            instance ?: synchronized(this) {
+ instance ?: HybridEdgeAgent(ctx.applicationContext).also { instance = it }
+            }
+    }
+}
+<span class="highlight">// MainActivity.kt - Uso</span>
+class MainActivity : AppCompatActivity() {
+    private val agent by lazy { HybridEdgeAgent.getInstance(this) }
+    override fun onCreate(savedInstanceState: Bundle?) {
+        super.onCreate(savedInstanceState)
+        lifecycleScope.launch {
+            agent.initModel("/sdcard/models/nemotron-3-nano-30b-q4.gguf")
+            val answer = agent.reason("Resume y traduce este documento...")
+            Log.d("Agent", answer)
+        }
+    }
+}
+</code></pre>
+    </div>
+    <div class="card">
+      <h2>⚡ Aceleración NumPy en Android (Chaquopy / Kivy)</h2>
+      <p>Si usas Python en Android con <strong>Chaquopy</strong> o integración JNI, compila NumPy contra <strong>OpenBLAS</strong> con soporte ARM NEON/SVE para aceleración vectorial.</p>
+      <pre><code># chaquopy build logic (build.gradle)
+python {
+    pip {
+        install "numpy"  <span class="highlight"># preferiblemente con wheel optimizado para ARM</span>
+        install "llama-cpp-python" <span class="highlight"># para inference ARM vía llama.cpp</span>
+    }
+}</code></pre>
+    </div>
+  </div>
+  <!-- AUTOMATIZACIÓN -->
+  <div id="auto" class="panel">
+    <div class="card">
+      <h2>🤖 Módulo de Automatización: Cuentas + Documentos + APIs</h2>
+      <button class="copy" onclick="copy(this)">Copiar</button>
+      <pre><code><span class="highlight"># automation_engine.py - Motor de automatización agéntica</span>
+import asyncio
+from playwright.async_api import async_playwright
+from dataclasses import dataclass
+from typing import Optionalimport base64, json
+@dataclass
+class AccountConfig:
+    service: str       <span class="highlight"># ej: "github", "aws", "huggingface"</span>
+    username: str
+    email: str
+    strategy: str = "llm_guided"  <span class="highlight"># llm_guided o deterministic</span>
+class AutomationEngine:
+    def __init__(self, nemotron_engine):
+        self.nemotron = nemotron_engine  <span class="highlight"># instancia de NemotronEngine</span>
+        self.browser = None
+        self.page = None
+    async def start(self, headless=False):
+        self.playwright = await async_playwright().start()
+        self.browser = await self.playwright.chromium.launch(headless=headless)
+        self.page = await self.browser.new_page()
+    async def read_document(self, url_or_path: str) -&gt; dict:
+        <span class="highlight">"""Lee PDF o página web, extrae texto, resume y traduce."""</span>
+        if url_or_path.endswith(".pdf"):
+            await self.page.goto(f"https://docs.google.com/viewer?url={url_or_path}")
+            content = await self.page.inner_text("#viewer")
+        else:
+            await self.page.goto(url_or_path)
+            content = await self.page.inner_text("body")
+        <span class="highlight"># Envía al LLM para análisis profundo</span>
+        analysis = self.nemotron.reason(
+            f"Resume, extrae pasos clave y traduce esta documentación:\n\n{content[:8000]}",
+            enable_thinking=True
+        )
+        return analysis
+    async def create_account(self, config: AccountConfig) -&gt; bool:
+        <span class="highlight">"""Automatiza flujo de registro GUI usando LLM para navegación adaptativa."""</span>
+        guide = self.nemotron.reason(
+            f"Genera un plan paso a paso para crear una cuenta en {config.service}. "
+            f"Incluye URLs, selectores CSS probables, validaciones de email y TOS.",
+            enable_thinking=True
+        )["answer"]
+        <span class="highlight"># Parsear plan y ejecutar acciones Playwright</span>
+        steps = self._parse_plan(guide)
+        for action in steps:
+            if action["type"] == "goto":
+                await self.page.goto(action["url"])
+            elif action["type"] == "fill":
+                await self.page.fill(action["selector"], action["value"].replace("{{email}}", config.email))
+            elif action["type"] == "click":
+                await self.page.click(action["selector"])
+            await asyncio.sleep(0.5)
+        return True
+    def _parse_plan(self, plan_text: str) -&gt; list:
+        <span class="highlight">"""Convierte respuesta textual de Nemotron a estructura ejecutable."""</span>
+        <span class="highlight"># Implementación con regex/JSON parsing robusto</span>
+        return [{"type":"goto","url":"https://github.com/signup"},{"type":"fill","selector":"#email","value":"{{email}}"}]
+    async def discover_api_and_setup(self, service_description: str):
+        <span class="highlight">"""Busca APIs vía HF Discover y configura tokens/env."""</span>
+        from orchestrator import HFAgentCLI
+        results = HFAgentCLI.search_skill(f"API for {service_description}", kind="mcp")
+        <span class="highlight"># El LLM selecciona la mejor API y genera código de setup</span>
+        choice = self.nemotron.reason(f"Selecciona la mejor API de este listado: {json.dumps(results)}")
+        return choice    async def close(self):
+        await self.browser.close()
+        await self.playwright.stop()
+<span class="highlight"># Ejecución de ejemplo</span>
+async def main():
+    agent = AutomationEngine(nemotron_engine=NemotronEngine("nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16"))
+    await agent.start(headless=False)
+    <span class="highlight"># 1. Leer documentación de una API desconocida</span>
+    doc = await agent.read_document("https://huggingface.co/docs/huggingface_hub/guides/cli")
+    print(doc["answer"])
+    <span class="highlight"># 2. Crear cuenta adaptativamente</span>
+    await agent.create_account(AccountConfig(service="huggingface", username="dev01", email="dev01@ai.dev"))
+    await agent.close()
+if __name__ == "__main__":
+    asyncio.run(main())
+</code></pre>
+    </div>
+    <div class="card">
+      <h2>🌐 Uso de HF Discover para Encontrar APIs Automáticamente</h2>
+      <pre><code># Descubre skills y MCP servers sin dejar la terminal
+hf discover search "image generation API" --kind mcp --json
+hf discover search "transcribe audio API" --kind skill --json
+hf discover search "train vision model" --limit 5
+<span class="highlight"># Instala el skill directamente para que el agente lo use</span>
+hf skills add --global</code></pre>
+    </div>
+  </div>
+  <!-- DEPLOY -->
+  <div id="deploy" class="panel">
+    <div class="card">
+      <h2>🚀 Guía de Despliegue Multi-Escenario</h2>
+      <table style="width:100%; border-collapse:collapse; color:#e5e7eb; font-size:.9rem;">
+        <tr style="background:#0e1525"><th style="padding:.6rem; text-align:left; border-bottom:1px solid #334155;">Escenario</th><th style="padding:.6rem; text-align:left; border-bottom:1px solid #334155;">Modelo</th><th style="padding:.6rem; text-align:left; border-bottom:1px solid #334155;">Runtime</th><th style="padding:.6rem; text-align:left; border-bottom:1px solid #334155;">Hardware</th></tr>
+        <tr><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Cloud GPU</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Nemotron-3-Ultra-550B-A55B</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">vLLM / TensorRT-LLM</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">2x B200 / 8x H100</td></tr>
+        <tr style="background:#0b0f19"><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Workstation</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Nemotron-3-Super-120B-A12B-FP8</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">TensorRT-LLM / vLLM</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">2x RTX Pro 6000 /2x H100</td></tr>
+        <tr><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Edge / Developer</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Nemotron-3-Nano-30B-A3B (NVFP4)</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">TensorRT-LLM / llama.cpp</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">DGX Spark / RTX 5090 / Jetson Thor</td></tr>
+        <tr style="background:#0b0f19"><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Android NPU</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Nemotron-3-Nano Q4/K quantized</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">MLLM + QNN / Cactus</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Snapdragon 8 Gen4 / Dimensity 9400</td></tr>
+        <tr><td style="padding:.6rem; border-bottom:1px solid #1f2937;">ARM Linux Edge</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">Nemotron-H-47B / Nano 30B</td><td style="padding:.6rem; border-bottom:1px solid #1f2937;">llama.cpp + KleidiAI / Arm NN</td><td style="padding:.6rem; border-bottom:1f2937;">Raspberry Pi 5 (Cortex-A76) / AGX Orin</td></tr>
+      </table>
+    </div>
+    <div class="grid-2">
+      <div class="card">
+        <h2>🐳 Dockerfile - Cloud Agent</h2>
+        <pre><code>FROM nvidia/cuda:12.6-devel-ubuntu24.04
+RUN pip install vllm huggingface_hub transformers playwrightRUN playwright install-deps
+RUN hf skills add --global
+ENV HF_TOKEN=${HF_TOKEN}
+CMD ["python", "orchestrator.py"]</code></pre>
+      </div>
+      <div class="card">
+        <h2>📦 Dockerfile - Edge ARM</h2>
+        <pre><code>FROM arm64v8/python:3.12-slim
+RUN pip install llama-cpp-python --extra-index-url ...
+COPY models/ /models/
+COPY orchestrator.py /
+CMD ["python", "orchestrator.py", "--edge"]</code></pre>
+      </div>
+    </div>
+    <div class="note">
+ 🚀 <strong>Quick start:</strong> Para probar el razonamiento sin GPU masiva, usa el modelo <code>nvidia/OpenReasoning-Nemotron-32B</code> (derivado de Qwen2.5) con vLLM en una sola H100 o incluso RTX 4090 con cuantización.
+    </div>
+  </div>
+</div>
+<script>
+  function show(id) {
+    document.querySelectorAll('.panel').forEach(p => p.classList.remove('active'));
+    document.querySelectorAll('.tab').forEach(t => t.classList.remove('active'));
+    document.getElementById(id).classList.add('active');
+    event.target.classList.add('active');
+  }
+  function copy(btn) {
+    const code = btn.parentElement.querySelector('pre').innerText;
+    navigator.clipboard.writeText(code);
+    btn.textContent = "¡Copiado!";
+    setTimeout(() => btn.textContent = "Copiar", 1500);
+  }
+</script>
+</body>
+</html>