Sujeeth Shetty

Support Fermix with any amount. It goes to the work below, in priority order: security, the fully local path, measured reliability, and signed releases.

Priority 1. Fermix runs commands and edits files on its owner's computer, so its safety boundaries have to hold. Funds an independent security review of the sandbox, signed updates and plugins, phone pairing and encryption, and the defenses against instructions hidden in web pages and messages, plus the fixes. Also brings the kernel-enforced command sandbox macOS already has to Linux (Landlock), so commands stay in the folders the owner allowed.

Priority 1. Model API credits set aside for a security review of every release before it ships: the agent, the Mac, iPhone, and Android apps, and the plugin catalog, each checked against the sandbox, signing, pairing, and encryption rules. Fermix ships about two releases a week, so a one-time outside review cannot keep up on its own; this keeps every release reviewed between independent audits. Kept separate from everyday development credits so it is never the budget that runs out.

Priority 2. Supports a Mac Studio with 512 GB of unified memory, enough to run the largest open-weight models entirely on the machine. This is how Fermix makes the fully local path real on macOS: local models through Ollama and on-device speech-to-text, so anyone can run Fermix with nothing leaving their computer and nothing to pay per message. A one-time purchase that replaces recurring cloud GPU costs.

Priority 2. Supports two NVIDIA DGX Spark machines (about $5,000 each) to build and test the same fully local path on Linux and NVIDIA GPUs, the setup most people self-hosting on Linux will have. Linked together they run larger models than one can alone, and they double as the arm64 Linux test machines for every release.

Priority 3. Supports model usage and cloud Linux machines (today on free tiers) for the Fermix capability benchmark, its private holdout (kept out of the public repo so no model can train on the answers), regression runs on every change, and rankings across models, so reliability claims are measured, not assumed.

Supports test phones (an iPhone and two Android phones from different makers) and a year of hosting for the two relays, in the US and the EU, that let a phone reach its owner's computer across networks. The relays carry only end-to-end encrypted traffic: no accounts, no stored messages, nothing they can read.

Supports metered API usage: the eight model providers Fermix works with (OpenAI, Anthropic, SpaceXAI, OpenRouter, Mistral, Venice, OpenAI Codex, and local Ollama) and the automatic fallback between them; realtime voice, voice-note transcription, WhatsApp Business, and paid plugin APIs, tested end to end on each release; and coding-agent costs for solo development and maintenance.

Supports Apple Developer membership (Mac and iPhone), Google Play registration, an Intel Mac, and CI time beyond today's free tier, so every release for macOS, Linux (Debian, Ubuntu, Fedora, openSUSE; x86_64 and arm64), iPhone, and Android is built, signed, and verified. Fermix keeps no user data on any server, so the rest stays small: the website, docs, install script, update feed, and domain, plus issue triage and community support.

Full funding for the current development phase, covering every plan above: an independently reviewed sandbox enforced by the operating system on macOS and Linux, a security review on every release, a fully local path on Mac and Linux, iPhone and Android apps that reach your own computer end-to-end encrypted, reliability measured across models, and signed releases everywhere. This request covers costs only; my development time is not included.