SHARVIN.OS boot sequence — stack, LLM routing, local inference, protocols and infra.
I architect and ship intelligent systems end to end — then run them in production for real users, on hardware I own.
Most of my work sits at one seam: natural language in, real infrastructure action out. A message becomes a tool call becomes a container restart becomes a reply. Getting that loop to be fast, cheap and reliable across five model providers is the interesting part.
chat/voice → intent + tool selection → service mesh → state change → grounded reply
↑ ↓
cheap fast models frontier models only
(routing, extraction) where reasoning earns it
Go · whatsmeow · MCP
A WhatsApp automation platform running 24/7. 20+ services, multi-model AI routing with automatic failover, media pipelines, MCP tool-calling and a Telegram bridge. Every inbound message is a tool-use loop.
IndicF5 · PyTorch · RTX 3090
Zero-shot voice cloning in Malayalam, served from a home GPU rig with real-time synthesis and a live waveform demo. Low-resource-language TTS that actually sounds like a person.
Unraid · Docker · Qdrant · OCI
A private cloud and AI cluster. GPU passthrough, local inference (Ollama / Qwen-class), vector + cache tier, Cloudflare tunnels and an OCI edge node for the things that must not go down.
◢ OPENCLAW-ON-PI · OPEN SOURCE
Python · Telegram · multi-provider
A production AI chatbot on a Raspberry Pi that routes between Groq, Ollama Cloud and local Ollama with automatic failover, rate limiting and a monitoring dashboard.
Also open:
webos-sidecar — local-first LG webOS sideload dashboard ·
HarmonyForge — media-library hygiene for Plex / Jellyfin / Emby ·
velvet-rows — language-aware Jellyfin home rows ·
govee-bt-proxy — Home Assistant Bluetooth proxy
Languages
Go TypeScript Python Kotlin C# AI / Inference
Anthropic Gemini Groq Ollama PyTorch MCP Web / Infra
Next.js React Docker Unraid Cloudflare Vercel
Website Photography Instagram Email
// SYSTEM ONLINE — IDENTITY CONFIRMED