AI agents and MCP servers LinkedIn GitHub @expertvagabond

Work

Shipped, Live, Clickable

My own products, the open-source tooling behind them, and websites built for paying clients. If it’s on this page, you can click it.

The Lab

Products I’m building myself, each at a different stage.

The Coldstar website: your USB drive is a cold wallet.

Launched · coldstar.dev

Coldstar

Any USB drive becomes an air-gapped signing device.

Coldstar boots a minimal Alpine Linux image from a USB stick, generates keys offline, and signs transactions on a machine that never touches the network. The Rust signer locks key material in memory and exposes it for about 100 microseconds before wiping it. On top sits a zero-knowledge layer I wrote from scratch.

  • 4× Superteam hackathon winner
  • Colosseum hackathon entry
  • Co-founded with @devsyrem
  • Solana
  • Rust · Ed25519 · ZK
The SINT Labs website.

Launched · sint.gg

SINT Labs

A voice agent that learned to trade, then grew into a control plane for AI.

SINT started at Kraken’s House of Ink residency as a voice-activated agent that my team and I taught to trade. It has since grown into an open-source protocol for governing AI agents and physical AI: the control plane that sits between autonomous systems and the real world.

  • Voice agent → trading → governance
  • Open source
  • Built at Kraken’s House of Ink
47%cheaper than the OpenAI API 180 msmedian latency 16×faster video encode

Building

AIDP: Distributed Inference

Serving LLMs on distributed GPUs for less.

A self-hosted LLM inference layer on decentralized GPU capacity: 47% lower serving cost than the OpenAI API at 180 ms median latency and 50 requests a second, with a guardrails gateway and evaluation datasets that score accuracy, safety and latency. A companion GPU video pipeline encodes 16× faster at a VMAF quality score of 95.8.

  • vLLM · llama.cpp
  • Guardrails + evals
  • GPU video pipeline

Open Source, Under My Own Name

The MCP servers and packages behind the products. All public, and you can check every one of them.

  • ZetaChain MCP

    The platform’s MCP server, built while I worked in developer relations there: the ZetaChain CLI, token and transaction tools exposed to AI assistants. Published to npm, then rewritten as a single Rust binary.

  • GOAT Network MCP

    91+ tools that give any AI agent JSON-RPC access to GOAT Network, a Bitcoin layer 2: blocks, transactions, balances, contract reads and logs.

  • Hedera MCP

    73 tools across Hedera’s account, token, consensus, EVM, file and scheduling services. It builds transactions and never holds keys.

  • Solana MCP

    Wallet management, transactions and program interactions for AI assistants, published as @expertvagabond/solana-mcp-server.

140Public repos
94npm packages
70Rust crates
10Hugging Face models

Before AI: Streaming at Scale

I architected and scaled a white-label real-time video backend to more than 50,000 concurrent users, for DirecTV LATAM, AT&T and IBM.

Custom HLS and FFmpeg pipelines, built while I worked in media and streaming, including at IBM Cloud Video (Ustream) on a product line worth more than $100M. That’s where I learned what production means: when it breaks, people notice immediately.

50K+Concurrent users
100+Cloudflare deployments

Let’s Build

Tell me what you’re building, what’s stuck, and when you need it. Every message comes straight to me, and I reply personally.

Prefer a DM? I’m on LinkedIn and X.

Enter a valid email address.
What’s this about?
Please add a message.