We need to output only the final README.md markdown, no preamble. Must be English, technical, precise, no fluff. Show real technical depth: LLM inference research (EAGLE/MTP/speculative decoding/MoE routing), running LLMs on unsupported consumer/datacenter GPUs (Tesla P40 sm_61, CUDA workarounds), self-built AI infrastructure, embedded systems.
Structure: centered header → 3-line technical bio → compact stack table → 4 project sections (LLM Research / AI Infrastructure / Tools & Automation / Embedded & Hardware) → GitHub stats.
For each project: [name](https://github.com/chessarisilvio/slug) + 1-line technical description.
Use repo slugs derived from project names (lowercase, hyphens).
Max 2 emojis total (header only), rest is clean technical prose.
90–120 lines total.
We need to base on vault data. We have a lot of project names. Need to pick representative projects for each section.
We need to produce a GitHub profile README.md for Silvio Chessari.
We need to include a centered header (maybe using HTML
). Use max 2 emojis in header only.
Then 3-line technical bio: short lines.
Then compact stack table: maybe a markdown table with columns: Language, Framework, Tools.
Then 4 project sections each with heading and list of projects.
Then GitHub stats: maybe using shields.io or just plain text.
We need to keep lines 90-120.
We must not include any extra text outside the markdown.
Let's craft.
First, header: maybe:
<div align="center">
# Silvio Chessari
</div>But need max 2 emojis in header only. Could add maybe a computer and a gear emoji? But we can avoid emojis to be safe. Use none.
But requirement: max 2 emojis total (header only). So we can have zero.
We'll do:
<div align="center>
# Silvio Chessari
</div>But need correct syntax: <div align="center">.
Now 3-line technical bio: maybe:
LLM inference researcher | AI systems builder | Embedded hardware engineer
Focus: speculative decoding, MoE routing, GPU kernel tuning for legacy NVIDIA architectures
Self-hosted LLM stack on Tesla P40 (sm_61) & RTX 3050, Arduino/R4 WiFi, ESP32 HID controllers
That's three lines.
Now compact stack table: maybe a table with two columns: Category and Technologies.
We'll do:
| Category | Technologies |
|----------|--------------|
| Languages | C++, Python, Rust, Bash |
| Frameworks | llama.cpp, TensorRT-LLM, PyTorch, ComfyUI |
| GPU/Accelerators | Tesla P40 (CUDA 11.8, sm_61), RTX 3050 (CUDA 12.x) |
| Embedded | Arduino R4 WiFi, ESP32, ESP-IDF, PlatformIO |
| Tools | Docker, systemd, MQTT/Tailscale, Obsidian vault, GitHub Actions |
That's maybe 6 rows.
Now sections:
List projects: from vault: "Add video input support to llama.cpp (mtmd)", "AI Gateway in prod: alternative concrete a LiteLLM", "Auto-Quantization Pipeline GGUF", "Nex2 Mini Phase Twin 30B low‑VRAM GGUF model", maybe "AI Model Selector Physical Controller" (though that's embedded). Also "AI Influencer" is more media.
We'll pick 4-5 projects per section? Probably each section list a few.
We need to output each as [name](https://github.com/chessarisilvio/slug) + 1-line technical description.
We need to derive slugs: lowercase, hyphens.
Examples:
- "Add video input support to llama.cpp (mtmd)" -> slug:
add-video-input-support-to-llamacpp-mtmd - "AI Gateway in prod: alternative concrete a LiteLLM" -> slug:
ai-gateway-in-prod-alternative-concrete-a-litellm - "Auto-Quantization Pipeline GGUF" -> slug:
auto-quantization-pipeline-gguf - "Nex2 Mini Phase Twin 30B low‑VRAM GGUF model" -> slug:
nex2-mini-phase-twin-30b-lowvram-gguf-model - "AI Model Selector Physical Controller" -> slug:
ai-model-selector-physical-controller - "AI Dashboard" -> slug:
ai-dashboard - "AI Home Assistant HID Dashboard" -> slug:
ai-home-assistant-hid-dashboard - "Voice Dictate — dettatura locale Whisper per Claude Code" -> slug:
voice-dictate-dettatura-locale-whisper-per-claude-codemaybe too long; we can shorten:voice-dictate-whisper-claude - "Auto Vault Journal" -> slug:
auto-vault-journal - "Blender 3D Generation" -> slug:
blender-3d-generation - "Barocco Digital" -> slug:
barocco-digital - "AI Influencer" -> slug:
ai-influencer
We need to keep lines count 90-120. Let's estimate.
We'll produce sections with maybe 3-4 projects each.
Now GitHub stats: maybe a line with shields or plain text: GitHub: @chessarisilvio | Public repos: 27 | Followers: 12 etc. But we don't have actual numbers; we can approximate or leave generic? Should be based on vault? Not provided. We can just put placeholder like GitHub: @chessarisilvio and maybe a line with 📊 but that's emoji; we can avoid emojis. Use plain text.
We need to keep total lines 90-120. Let's draft and count.
We'll write markdown.
Let's start constructing.
Line count includes empty lines? Usually yes. We'll aim for about 100 lines.
We'll write:
<div align="center">
# Silvio Chessari
</div>
LLM inference researcher | AI systems builder | Embedded hardware engineer
Focus: speculative decoding, MoE routing, GPU kernel tuning for legacy NVIDIA architectures
Self-hosted LLM stack on Tesla P40 (sm_61) & RTX 3050, Arduino/R4 WiFi, ESP32 HID controllers
| Category | Technologies |
|----------|--------------|
| Languages | C++, Python, Rust, Bash |
| Frameworks | llama.cpp, TensorRT-LLM, PyTorch, ComfyUI |
| GPU/Accelerators | Tesla P40 (CUDA 11.8, sm_61), RTX 3050 (CUDA 12.x) |
| Embedded | Arduino R4 WiFi, ESP32, ESP-IDF, PlatformIO |
| Tools | Docker, systemd, MQTT/Tailscale, Obsidian vault, GitHub Actions |
## LLM Research
- [add-video-input-support-to-llamacpp-mtmd](https://github.com/chessarisilvio/add-video-input-support-to-llamacpp-mtmd) – Enables real‑time video frame ingestion for llama.cpp inference pipelines.
- [ai-gateway-in-prod-alternative-concrete-a-litellm](https://github.com/chessarisilvio/ai-gateway-in-prod-alternative-concrete-a-litellm) – Comparative study of LiteLLM alternatives for local OpenAI‑compatible routing on P40/3050.
- [auto-quantization-pipeline-gguf](https://github.com/chessarisilvio/auto-quantization-pipeline-gguf) – Automated GGUF quantization benchmarking Q4_K_M vs Q5_K_S on Tesla P40.
- [nex2-mini-phase-twin-30b-lowvram-gguf-model](https://github.com/chessarisilvio/nex2-mini-phase-twin-30b-lowvram-gguf-model) – Low‑VRAM GGUF port of Nex2 Mini Phase Twin 30B for 24 GB P40.
## AI Infrastructure
- [ai-dashboard](https://github.com/chessarisilvio/ai-dashboard) – Web UI at :9190 monitoring GPU, services, agenda and security scans.
- [ai-home-assistant-hid-dashboard](https://github.com/chessarisilvio/ai-home-assistant-hid-dashboard) – Arduino R4 WiFi + ESP32 HID dashboard showing VRAM, tok/s, uptime via MQTT/Tailscale.
- [auto-vault-journal](https://github.com/chessarisilvio/auto-vault-journal) – Hook‑driven Obsidian vault updater and session journal for Claude Code.
- [voice-dictate-whisper-claude](https://github.com/chessarisilvio/voice-dictate-whisper-claude) – Space‑bar Whisper turbo GPU dictation plug‑in for Claude Code terminal.
## Tools & Automation
- [blender-3d-generation](https://github.com/chessarisilvio/blender-3d-generation) – Procedural asset pipeline on RTX 3050 using CC0 models and HDRI lighting.
- [barocco-digital](https://github.com/chessarisilvio/barocco-digital) – Astro+Tailwind v4 agency site with Cloudflare Pages deploy and accessibility audit.
- [ai-influencer](https://github.com/chessarisilvio/ai-influencer) – LoRA‑tuned SDXL pipeline for consistent face generation (character “Aria”).
- [ai-model-selector-physical-controller](https://github.com/chessarisilvio/ai-model-selector-physical-controller) – ESP32 rotary encoder + OLED HID controller for local model switching.
## Embedded & Hardware
- [ai-home-assistant-hid-dashboard](https://github.com/chessarisilvio/ai-home-assistant-hid-dashboard) – Duplicate? Already listed; maybe move to Embedded & Hardware only. Let's adjust: we can have it only in Embedded & Hardware, and remove from AI Infrastructure. Let's restructure.
We need 4 sections: LLM Research, AI Infrastructure, Tools & Automation, Embedded & Hardware.
We'll move the dashboard to Embedded & Hardware.
Let's recraft sections.
### LLM Research (as above)
### AI Infrastructure
- [ai-dashboard](https://github.com/chessarisilvio/ai-dashboard) – Web UI at :9190 monitoring GPU, services, agenda and security scans.
- [auto-vault-journal](https://github.com/chessarisilvio/auto-vault-journal) – Hook‑driven Obsidian vault updater and session journal for Claude Code.
- [voice-dictate-whisper-claude](https://github.com/chessarisilvio/voice-dictate-whisper-claude) – Space‑bar Whisper turbo GPU dictation plug‑in for Claude Code terminal.
- [ai-gateway-in