Welcome to Artificial Humanity! We are a research and engineering studio dedicated to building AI projects that touch upon the space between machine intelligence and human emotion.
Our core philosophy is simple: we believe artificial intelligence should run locally, respect privacy, and engage with human feeling. No remote cloud APIs, no data locks, and no delay. Just native, high-performance engines that feel alive.
Our flagship initiative is Project Prosodia —an open-source framework for expressive, continuous, emotion-driven speech synthesis. It splits speech generation into a protocol-based pipeline, separating semantic interpretation (the Director) from vocal performance (the Actor, powered by our own Sonora model), so digital books are read with dynamic pacing, dramatic pauses, and real-time voice blending—no robotic conveyor-belt narration here.
Fun facts: we're hunting for the exact Valence-Arousal-Tension coordinate for "the hero's tragic sacrifice," to get the Actor's voice to crack slightly. And since the engine compiles and runs entirely offline, you'd still have a perfect narrator even stranded on a mountaintop with a laptop.
The voice inside Prosodia is its own project: Sonora —a directable, castable, mobile-friendly TTS actor we train from scratch on license-clean data. It is built on the Matcha-TTS architecture (a conditional flow-matching mel decoder solved with a few-step ODE), augmented with emotion conditioning and continuous voice casting, so valence, energy and tension arrive as things you tell the model rather than hope for. Small enough to run on a phone, and Apache-2.0 like everything else here.
Another studio initiative is Council of Experts —a native macOS orchestration platform that runs a configurable council of LLM experts (Claude, Gemini, GPT, Grok, and local models) in parallel, has them critique and revise each other's drafts, and synthesizes their consensus through a Chairman model. It's built on the same philosophy as Prosodia: a Rust core with a native Swift/SwiftUI FFI layer, supporting live streaming, workspace file context, and multi-turn session persistence.
Lucida is the studio's own image and video generation tool, and a single Rust binary either way. It reaches Google Gemini and Veo, Runway, Kling, hosted FLUX, Stability AI, OpenAI, or a ComfyUI running on your own hardware. It works as a command-line tool and as an MCP server, so the coding agents working on these projects can make their own artwork instead of waiting for someone to hand it over.
FerroStep: a stepped path from a hollow initial-state node to a glowing terminal-state node
Also from the studio, FerroStep —a data-driven state-machine referee for multi-agent loops. Instead of hiding orchestration state inside a framework's graph object, FerroStep treats your own database as the single source of truth: AI agents and humans act as peers on one ledger, while a pure Rust core rules on every transition—who may move what, what a retry costs, and when a human takes over. Loop ceilings survive crashes, escalation is a routing decision rather than an error, and the whole system is inspectable with a database browser. True to the studio philosophy: an engine, not a platform—no runtime, no scheduler, no hosted anything.
FerroTrack is FerroStep's sibling—a light issue tracker built for the way agent work actually runs, where people and agents file, claim and close against one shared ledger rather than a chat log nobody can query afterwards. It is meant to be the store FerroStep reaches for out of the box, and a tracker in its own right. Early days, and built in the open.
Thanks for stopping by, and happy coding! 👤✨