Our products

• AI (service page)License: Apache 2.0

  GhostSwitch

  Local AI assistant

  One open Qwen3.5 model plus a library of LoRA experts for code, finance, law, CAD and two dozen other fields. For every message it picks the expert that fits, writes the answer and checks it in code. Runs on a laptop with 16 GB of memory.

  • 88.3%HumanEval+ with the coding expert, against 76.7% for the base model
  • 16 GBof memory runs the whole expert library
  • 27experts for Qwen3.5-9B

  What it does

  • A router picks the expert for each message, or answers with the base model when none fits
  • The expert attaches to the model already in memory in 0.1–0.3 s, with no reload
  • Answers are checked by code: completeness, arithmetic, citations, JSON fields; a failed check means a retry
  • OpenAI-compatible API: works with the OpenAI SDKs and the OpenCode coding agent
  • Chats stay on your computer; it works offline once the models are downloaded

  • Python
  • Qwen3.5
  • LoRA
  • MLX
  • llama.cpp
  • OpenAI API

  Code on GitHub (opens in a new tab)
• Web3 (service page)License: MIT

  Solana Shred Receiver

  Live Solana transactions

  Joins Solana mainnet Gossip as an unstaked node, receives the shreds Turbine sends it and rebuilds transactions from them; missing shreds are recovered with Reed-Solomon. Live transactions with no validator, no stake and no paid shred feed.

  • ~40 mson average from a received shred to a decoded transaction
  • $0for the shred feed: no paid provider behind it
  • 4 GBof RAM and 4 vCPU run one instance

  What it does

  • Decoded transactions without Jito ShredStream, DoubleZero or a paid RPC
  • Two decode paths: complete data shreds go out at once, missing ones are rebuilt from coding shreds
  • Duplicate shreds from different retransmitters are dropped before recovery
  • WebSocket streams of shreds and transactions, Prometheus metrics, JSON logs

  • Rust
  • Solana
  • Agave
  • Gossip
  • Turbine
  • Reed-Solomon
  • Prometheus

  Code on GitHub (opens in a new tab)
• Marketing automation (service page)

  RelayCore

  App events to Telegram

  A serverless relay on Vercel: your apps send events to one HTTP endpoint, RelayCore delivers them to Telegram bots and chats. Bots, routes and chats are set up in a web dashboard, with no server to run.

  • 1HTTP endpoint for events from all your apps
  • 0servers to manage: it deploys to Vercel
  • AES-256encryption of bot tokens and API keys

  What it does

  • Server errors, deploys, payments and cron jobs arrive as Telegram messages
  • One event goes to many chats at once, including topics in forum groups
  • A live event log with the delivery status of every message
  • Bot tokens and API keys encrypted with AES-256-GCM; a setup wizard on first run

  • Next.js 16
  • TypeScript
  • Upstash Redis
  • Telegram Bot API
  • Vercel
  • Tailwind CSS

  Code on GitHub (opens in a new tab)

All numbers are from each project’s README on GitHub.