Founder & Lead Architect @ SMLabs AI · Building AI systems that run anywhere — cloud, edge, or offline
I'm an AI Systems Architect focused on the orchestration primitives, cognitive state machines, and inter-agent protocols that make agentic systems reliable at runtime.
smagent / Cognitive Runtime (private) Agentic OS with real orchestration primitives, cognitive state management, and inter-agent communication protocols. Built from scratch for full production control — no framework lock-in.
anthropic-proxy (public) Multi-provider LLM routing layer with protocol translation — routes Claude Code CLI through Gemini, Groq, DeepSeek, and OpenRouter. Direct provider integration — minimal abstraction overhead.
MedGemma FORENSIC (public) Offline multimodal AI diagnostics on a $200 Android device. Gemma 3 1B + MedSigLIP + MedASR at 417MB peak RAM. Submitted to Google's Health AI Developer Foundations program.
SMExpense (private) AI-powered expense analyzer with receipt scanning, voice input, and smart budgeting.
Serper Search MCP Server 8-tool MCP server (web, images, videos, news, shopping, places, deep research, RAG context) used with Claude, Cursor, Windsurf, n8n, and KiloCode. Published on npm and Docker. TypeScript, full type safety, zero telemetry.
Running a Multimodal Diagnostic Suite (LLM + Vision + Audio + Anomaly Detection) on a consumer Android device under 500MB peak RAM.
A "Traffic Cop" sequential AI kernel (ModelLifecycleManager.kt) orchestrates ~4.3GB of model weights serially — keeping peak native heap at ~417MB on a 4GB RAM device.
Worker Node (Field Medic) Anchor Node (Command Center)
├── MedSigLIP → visual capture ├── Receives CaseFile.proto streams
├── MedASR → verbal autopsy ├── Full diagnostic reasoning
└── Packages → CaseFile.proto ├── Gemma 3 1B → ForensicReport
↓ P2P Mesh / WiFi Direct ↓ └── AnomalyEngine → Epidemic Clusters
AI / ML & Agents
-FF6F00?style=flat-square&logo=tensorflow&logoColor=white){kind=icon}
LLM Providers & APIs
Full Stack SaaS
Edge & Mobile
Backend & APIs
Infrastructure & DevOps
Self-hosted on bare-metal Linux — no managed platform abstraction layer.
netcup VPS 1000 ARM G11 · Manassas, USA
├── 🔲 6-core ARM64 · 8GB RAM · 256GB NVMe SSD
├── 🐧 Ubuntu 24.04 LTS
├── 🐳 Docker · Nginx · Traefik reverse proxy · SSL
├── 🚀 Coolify — self-hosted deployment orchestration
└── 🧠 Hosts: smagent Cognitive Runtime · future SMLabs AI services
Provisioning, securing, and operating Linux servers without a managed platform layer.
| System | What It Is | Stack |
|---|---|---|
| medgemma-forensic (public) | Edge AI pathologist — multimodal diagnostics on Android | Kotlin, LiteRT, ONNX, Protobuf, P2P |
| serper-search-mcp-server (public) | Enterprise MCP server for agent web search | TypeScript, MCP, Node.js, Serper API |
| smagent (private) | Cognitive Runtime — agentic OS, orchestration primitives | Python, LangGraph, Redis, MCP |
| anthropic-proxy (public) | LLM gateway routing Claude Code through Gemini, Groq, DeepSeek | Python, Claude API, Gemini, Groq, DeepSeek |
| smexpense (private) | Live AI expense SaaS — receipt OCR, voice input, smart budgeting | Next.js, Supabase, Prisma, Tailwind, PostgreSQL, Redis |
"The difference between an AI enthusiast and an AI architect is orchestration."