• automated-exl2-conversion-validation-pipeline C++/Python toolchain for Qwen3_5MoeForC architecture fixes and mixed-precision EXL2 conversion.

  • add-video-input-support-to-llamacpp-mtmd CMake and C++ patches enabling webcam/video frame capture for multimodal llama.cpp inference.

  (Section 2: AI Infrastructure)

  🖥️ AI Infrastructure & Gateway Systems

  • ai-gateway-in-prod-alternative-concrete-a-litellm Analysis and deployment of OpenAI-compatible routing gateways optimized for 24GB/8GB VRAM constraints.
  • openclaw Node.js Ollama gateway and local AI service orchestrator.
  • ai-dashboard Real-time monitoring dashboard (:9190) tracking GPU metrics, token throughput, and automated AGENDA.md generation.
  • secure-llm-context-vault Encrypted local storage architecture for managing LLM context windows and prompt templates.

  (Section 3: Tools & Automation)

  🛠️ Tools & Automation

  • auto-vault-journal Obsidian vault synchronization system triggered by Claude Code Stop hooks with async transcript parsing.
  • voice-dictate Local Whisper turbo GPU integration for terminal-based voice input, bypassing native voice mode limitations.
  • reddit-monitor Automated AI/tech subreddit scraper feeding structured ideas into AGENDA.md via 3h systemd timers.
  • automazione-bootwatchdog-ai-avanzato Systemd-based watchdog monitoring llama-stack process health, VRAM allocation, and tok/s rates with Telegram alerts.

  (Section 4: Embedded & Hardware)

  📡 Embedded & Hardware Integration

  • ai-home-assistant-hid-dashboard Arduino R4 WiFi + ESP32 HID controller displaying P40/3050 telemetry via MQTT/Tailscale.
  • controller-termico-proattivo-esp32 Proactive thermal management firmware with dynamic fan curve adjustment and sensor fusion.
  • digital-thermal-lcd HID-based LCD display driver for real-time thermalright temperature monitoring.
  • ai-model-selector-physical-controller ESP32 rotary encoder interface for local model switching with OLED feedback and HID command injection.

  (GitHub Stats)

  

4. Check Constraints:

• Output ONLY final README.md? Yes.
• English, technical, precise, no fluff? Yes.
• Show real technical depth (EAGLE/MTP, P40 sm_61, etc.)? Yes, integrated into bio and stack.
• Structure: centered header → 3-line bio → compact stack table → 4 sections → stats? Yes.
• Projects: [name](url) + 1-line desc? Yes.
• Slugs derived from names? Yes.
• Max 2 emojis total (header only)? I used 🔬 and ⚙️ in the header. I need to remove any other emojis. I used 🔍, 🖥️, 🛠️, 📡 in section headers. I must remove those to strictly follow "Max 2 emojis total (header only)".
• 90-120 lines total? Let's count. Header: 3 Bio: 3 Table: 6 Section 1: 5 Section 2: 5 Section 3: 5 Section 4: 5 Stats: 3 Total ~35 lines. I need to expand slightly to hit 90-120, or add more projects/details while keeping it tight. I'll add a few more vault projects to fill it out and ensure line count is in range. I'll also format it carefully. Let's count lines precisely and adjust to hit 90-120. I will add more projects from the vault to each section to reach ~95 lines. Vault projects available: sistema-di-benchmarking-automatizzato-per-nuovi-mo, nex2-mini-phase-twin-30b-lowvram-gguf-model, bot-short, megatool, modulo-offline-exifgps-geotagging-per-megatool, videostudio, web-access-layer-per-agenti-ai-locali, automazione-systemd-timer-per-html-minification-gi. I'll distribute them across sections. Let's draft carefully and count lines.

Revised Draft (Mental Line Count Check):

AI infrastructure engineer specializing in local LLM inference, speculative decoding, and heterogeneous GPU acceleration. Architect of self-hosted AI stacks running quantized MoE and dense models on unsupported datacenter silicon (Tesla P40 sm_61) and consumer RTX hardware. Embedded systems developer bridging microcontrollers, MQTT, and real-time telemetry for autonomous AI workstations.