⚡ SERVER INFRASTRUCTURE MANAGEMENT

AI Supervisor
Total server control

Painel técnico de monitorização e gestão de infraestrutura de IA — Docker, GPU, Serviços, Rede, Backups, Cameras e Alerts num único dashboard em tempo real, construído com FastAPI + React + WebSockets.

FastAPI React 18 WebSockets Docker SDK NVIDIA GPU JWT Auth Ollama LLM psutil OpenCV QNAP NAS

Active modules

WS live interval

30+

API endpoints

AI Services

JWT

Auth HS256

🏗️ Architecture Técnica

Asynchronous full-stack

React 18 + Vite frontend communicates with FastAPI (Python) backend via REST and WebSocket. The backend aggregates data from Docker SDK, psutil, nvidia-smi and systemctl in real time.

FRONTEND — React 18 + Vite
├── React Router 6 // 15 rotas
├── Axios // automatic JWT interceptor
├── Recharts // time-series charts
├── Tailwind CSS // dark theme vars
└── WS nativo // live metrics at 3s

BACKEND — FastAPI (Python 3)
├── Uvicorn // ASGI server :9100
├── Docker SDK // Unix socket
├── psutil // CPU/RAM/Disk
├── pynvml // GPU VRAM/temp/util
├── OpenCV + face_rec // câmeras
├── bcrypt + PyJWT // auth HS256 24h
├── smtplib + SMS API // alertas
└── CIFS + rsync // backup QNAP

STORAGE — File-based
├── users.json // bcrypt, lockout, flags
├── secret.key // 48-byte hex random
└── *.json configs // alerts, cameras, VRAM

🔐 Login — JWT + bcrypt + rate limiting

24h

Token expiry

Max attempts / 15min

30min

Lockout after 10 failures

📊 Main Dashboard

Real-time view via WebSocket

Upon entering, the dashboard aggregates CPU, RAM, Disk, GPU metrics, AI service status and Docker containers — automatically updated every 3 seconds via WebSocket ws://…/ws/live.

Technical narration: "The AI Supervisor central dashboard presents all critical server indicators in real time: CPU and RAM usage with 60-sample history, NVIDIA GPU status with VRAM/temperature/utilization, and the service health grid aggregating AI Services, Docker Containers, open ports and active alerts. Data arrives via WebSocket broadcast at 3 seconds — no polling."

📊 Dashboard — live metrics, service health, Recharts charts

WebSocket broadcast interval

Ring buffer samples (CPU/RAM/GPU)

Top KPI cards (CPU/RAM/Disk/GPU)

Service health tiles

⚡

WebSocket Live Feed

Endpoint WS /ws/live broadcasts to all connected clients every 3 seconds with CPU %, RAM %, GPU util and VRAM.

📈

Ring Buffer Metrics

Backend maintains 60 in-memory samples for CPU, RAM, GPU utilization and VRAM %. Visualized with Recharts line charts.

🩺

Service Health Grid

Grelha 2×3 com status de AI Services, Docker Containers, System Services, GPU, Network Ports e Alerts — com contadores live.

🖥️

Top KPI Cards

CPU (% + cores + freq), RAM (GB used/total + swap), Disk (GB + %), GPU (VRAM + temp + util) — data via psutil + pynvml.

🧠 AI Services

Control of 8 AI services

Centralized management of all AI services running on the server — Ollama, Whisper, Piper TTS, Edge-TTS, XTTS, Emotion Engine, Wakeword and AI Router — with start/stop/restart and real-time logs.

Technical narration: "The AI Services module automatically discovers the 8 configured AI services, checks their status via HTTP healthcheck on each port, and allows start/stop/restart via systemctl or Docker. The administrator sees each service's uptime, the port it's listening on, and can access the last session's logs without leaving the panel."

🧠 AI Services — healthcheck + start/stop/logs

          Registered AI Services:

          ├── Ollama       :11434  LLM inference

          ├── Whisper STT  :10300  speech-to-text

          ├── Piper TTS    :10200  offline TTS

          ├── Edge-TTS     :10201  Azure neural

          ├── XTTS         :8020   advanced TTS

          ├── Emotion      :5055   sentiment analysis

          ├── Wakeword     :10400  voice activation

          └── AI Router    :8090   unified API GW

GET /api/services/ai POST /start|stop|restart GET /logs

🐳 Docker Management

Container management via Docker SDK

Lists, controls and monitors all running Docker containers — with access to logs, CPU/memory stats per container, and bulk operation to start all stacks at once.

Technical narration: "The Docker integration uses the Docker SDK for Python via Unix socket for direct communication with the Docker daemon. The panel lists all containers with status, image, exposed ports and the source docker-compose directory. Each container has start/stop/restart actions and log access. The 'Start All Stacks' button executes the ~/start_all_docker_stacks.sh script to bring up the entire infrastructure at once."

🐳 Docker Containers — status, logs, start/stop/restart, bulk ops

📋

Container List

Full list with status (running/stopped), image, ports, labels and source compose directory via Docker SDK.

📝

Real-Time Logs

Endpoint GET /api/docker/containers/{name}/logs with configurable number of lines via journalctl.

📊

Container Stats

CPU % and memory per container via GET /api/docker/containers/{name}/stats — direct data from the Docker daemon.

🚀

Start All Stacks

Bulk operation that executes the entire infrastructure startup script — useful after server reboot.

🎮 GPU & VRAM Management

NVIDIA GPU monitoring and automatic VRAM management

Complete NVIDIA GPU monitoring via pynvml — utilization, VRAM, temperature, power draw, fan speed and P-State. The VRAM Manager automatically unloads idle AI services to free up memory.

Technical narration: "The GPU module uses the pynvml library for direct reading of NVIDIA counters without relying on nvidia-smi subprocess. The VRAM Manager is an async watchdog running every 30 seconds: it compares each service's last activity timestamp with the configured timeout and, if exceeded, sends a stop to the service. The result is a dynamic VRAM pool that reconfigures itself based on services in use."

🎮 GPU Monitor — util/VRAM/temp/power/pstate

⚡ VRAM Manager — auto-unload idle AI services

30s

Watchdog check interval

2min

Idle timeout (Emotion/Whisper/XTTS)

5min

Idle timeout (Wakeword)

pynvml

GPU data source

      VRAM Manager — per-service configuration:

      Emotion Engine  idle_timeout=120s  auto_stop=true   // unloads if idle

      Whisper STT     idle_timeout=120s  auto_stop=true   // frees VRAM after use

      XTTS            idle_timeout=120s  auto_stop=true   // large model, frees up

      Ollama          idle_timeout=  —    auto_stop=false  // manual control

      Wakeword        idle_timeout=300s  auto_stop=false  // disabled by default

🔧 Infrastructure

Disk, network and system services

Disk analysis with automatic cleanup, network interface and open port management, full systemd service control with journalctl and aggregated server log visualization.

💾 Disk Manager — analysis + cleanup + migration

🌐 Network — interfaces + open ports

⚙️ System Services — systemctl + journalctl

📈 Metrics — time-series CPU/RAM/GPU charts

📄 Logs — system log aggregation

💾

Disk Manager

Partition analysis, Docker image/volume sizes, dangling image cleanup, large file scanner and volume migration.

🌐

Network & Ports

Lists network interfaces with statistics, open ports with PIDs and service names, and port availability checker.

⚙️

System Services

Systemd service control (start/stop/restart/enable/disable) with real-time logs via journalctl.

📈

Metrics History

60 CPU, RAM and GPU samples in ring buffers visualized as time-series charts with Recharts. Data via GET /api/metrics.

🗄️ Backup & Restore

Automatic backup to QNAP NAS via CIFS

Full backups of services, MariaDB databases, Docker volumes and data folders to QNAP NAS via CIFS/rsync — with selective restore per service or database.

Technical narration: "The Backup module mounts CIFS shares from the QNAP NAS and runs rsync to copy configured service data, runs mysqldump for each MariaDB database, and archives docker-compose YAMLs. Backups are catalogued in JSON with size and date metadata. Restore is selective — you can restore a single database or just one service's data without affecting the rest."

🗄️ Backup & Restore — QNAP NAS + rsync + mysqldump

          Services with backup configured:

          nextcloud · vaultwarden · bookstack

          n8n · taiga · wordpress

          nginx-proxy · mariadb

          wireguard · portfolio · mmia

          Backup targets:

          Manual:  /mnt/qnap_backup_manual

          Auto:    /mnt/qnap_backup_auto

          What is included:

          ✓ Docker compose YAMLs

          ✓ Data folders (/srv/dev-disk-*/data)

          ✓ MariaDB dumps (mysqldump)

          ✓ Docker root configuration

📹 Vigilância & Cameras

IP surveillance with face recognition and motion detection

Video surveillance system supporting RTSP/HTTP cameras, motion detection via OpenCV, facial recognition with face_recognition + dlib, and automatic email notifications.

Narração técnica: "O módulo de câmeras usa OpenCV para capturar frames de streams RTSP ou HTTP. A detecção de movimento compara frames consecutivos por limiar de pixels configurável. O reconhecimento facial usa a biblioteca face_recognition para comparar encodings com a base de dados de rostos conhecidos. Quando deteta um rosto desconhecido ou movimento, dispara notificação por email com snapshot. Cameras Android podem auto-registar-se via API."

📹 Camera Surveillance — RTSP/HTTP, motion, face recognition, events

🎥

MJPEG Live Stream

MJPEG proxy via GET /api/cameras/{id}/stream for any RTSP/HTTP camera — no browser plugins needed.

🏃

Motion Detection

OpenCV frame diff with configurable pixel threshold. Records event with timestamp and thumbnail. Notifies by email if active.

👤

Face Recognition

Known faces database with dlib encodings. Distinguishes "known face", "unknown face" and "no face". Separate alert per type.

🤖

AI Activity Analysis

Activity analysis via Ollama vision models — describes what is happening on the camera in natural language.

📱

Android Camera

Android phones register themselves as camera sources via POST /api/android/register.

🔍

Auto-Discovery

Automatic local network scan to discover IP cameras with brand detection and default credential suggestions.

🔔 Sistema de Alerts

Alerts por thresholds com recomendações de IA

Configurable thresholds per metric (CPU, RAM, GPU, VRAM, Disk, Temperature) with two levels (Warning/Critical), SMTP email and SMS Gateway notification, anti-spam cooldown and local LLM analysis.

🔔 Settings & Alerts — SMTP, SMS, thresholds, AI recs

          Configurable thresholds:

          cpu_percent     warn=75%  crit=90%

          ram_percent     warn=80%  crit=95%

          gpu_util        warn=85%  crit=95%

          gpu_temp        warn=75°C crit=85°C

          vram_percent    warn=80%  crit=90%

          disk_percent    warn=75%  crit=90%

          Notification channels:

          ✓ Email SMTP (HTML formatted)

          ✓ SMS Gateway API

          ✓ Ollama AI recommendations

          Anti-spam:

          configurable cooldown (min) + 200 alert history

🔐 Security & Autenticação

JWT + bcrypt + rate limiting + audit log

Production-ready authentication system with JWT HS256, bcrypt 12 rounds, IP rate limiting, automatic lockout for repeated failures, and complete login audit log.

Technical narration: "Security was designed to withstand brute force attacks: an IP exceeding 5 failed attempts in 15 minutes receives HTTP 429; after 10 consecutive failures the account is locked for 30 minutes. JWT uses a 48-byte randomly generated key on first startup, stored with 0o600 permissions. The audit log stores the last 200 login events with IP, timestamp and result — accessible only to superusers."

🔑

JWT HS256 — 24h

Token generated with 48-byte random key (secret.key, mode 0o600). Expires in 24h. Axios automatically injects into all headers.

🔒

bcrypt 12 rounds

Passwords stored with bcrypt (12 rounds) in users.json with 0o600 permissions. Never stored in plain text.

🛡️

IP Rate Limiting

Maximum 5 failed attempts per IP in a 15-minute window. Exceeded → HTTP 429. Implemented without external middleware.

🔓

Account Lockout

10 consecutive failures lock the account for 30 minutes. State saved in users.json with unlock timestamp.

📋

Audit Log

Last 200 login events in memory (IP, timestamp, result, user-agent). Accessible via GET /api/auth/login-audit — superusers only.

🔄

Force Password Change

must_change_password flag in user JSON. Modal blocks the app until the user sets a new password via POST /api/auth/change-password.

🚀 Launcher & Presentation

Application launching and integrated demo

Quick launch panel for configured applications and scripts, plus an internal presentation page with system information and platform version.

🚀 Launcher — quick launch for apps and scripts

Sidebar — 15 Modules

          Dashboard · Lançamentos · AI Services

          Containers · Disk Manager · Backup

          GPU Monitor · Network · System Services

          VRAM Manager · Metrics · Logs

          Settings · Cameras · Info

          Footer: MagWeb · v{version}

React Router 6 PrivateRoute wrapper Collapsible mobile GET /api/version

⚡ AI SUPERVISOR — COMPLETE STACK

Infrastructure de IA
sob controlo total

FastAPI + React + Docker SDK + NVIDIA pynvml + OpenCV + Ollama + QNAP CIFS — a single panel to manage the entire server AI infrastructure in real time.

⚡

WebSocket Live

Metrics push at 3s to all connected clients via WS /ws/live.

🧠

8 AI Services

Ollama, Whisper, Piper, Edge-TTS, XTTS, Emotion, Wakeword, AI Router.

🎮

VRAM Manager

Async watchdog unloads idle services — dynamic VRAM pool.

🐳

Docker SDK

Direct communication via Unix socket — no CLI invocation.

🗄️

QNAP Backup

CIFS + rsync + mysqldump to NAS with selective restore per service.

📹

Face Recognition

OpenCV + dlib for motion detection and facial recognition.

🔐

JWT + bcrypt

HS256 24h, 12 rounds, IP rate limiting, 30min lockout.

🔔

Alerts IA

Thresholds + SMTP + SMS + Ollama LLM recommendations.

FastAPI 0.110 React 18.3 TypeScript 5.4 Python 3 psutil 5.9 pynvml 11.5 Docker SDK 7.0 bcrypt + PyJWT OpenCV 4.9 Vite 5.3 Recharts 2.12 Tailwind 3.4 Uvicorn 0.29

AI Supervisor Total server control