AI that amplifies human capability.
Augmentation, not replacement. We build software that extends what a person can do — speak, or ship — on models we train and infrastructure we run ourselves.
What we build
Models we train ourselves
We fine-tune speech recognition models and serve speech synthesis on GPUs we operate, rather than wrapping someone else's endpoint in a new interface.
Products on every surface
Web applications, native iOS apps, Python services, and cross-platform desktop software. We build wherever the person we are building for already works.
Made to stay running
Authentication, storage, and continuous integration across macOS, Windows, and Linux behind what we ship.
Claudia
Run a fleet of coding agents from one screen, instead of one agent in one terminal.
Claudia keeps many agent sessions going at once across different projects and puts them all in front of you — live terminals, the diff each agent produced, and isolated git worktrees so parallel work never collides. Agents can create and coordinate their own sub-tasks, which changes the job from typing to directing.
- Many agent sessions running in parallel
- An isolated git worktree for each task
- Live terminals, per-task diffs, and revert
- An MCP server, so agents coordinate their own sub-tasks
- Works with Claude Code and OpenCode
- Desktop app for macOS, Windows, and Linux
npm install -g @extropolis/claudia NeuralSpeaker
Speech recognition that adapts to one person, for people whose speech general-purpose models fail on.
Most speech recognition is trained on the average speaker, which is precisely why it breaks down on atypical speech. NeuralSpeaker builds a model around the individual instead: someone records their own voice, corrects the machine's first attempt at transcribing it, and that human-verified dataset is used to fine-tune a Whisper model that belongs to them.
It can speak, too. Text is synthesised in the person's own voice style from a short reference sample using zero-shot synthesis — no voice model is created or retained — and every generated clip carries an inaudible provenance watermark marking it as machine-made.
- A speech recognition model fine-tuned per person
- Human-verified datasets, built by the user
- Zero-shot speech synthesis in the user's voice style
- Web app and a native iOS app
NeuralSpeaker is in private beta with a small group of users in the United States while we finish the compliance work that handling voice data properly requires.
How we work
We are a small team. We treat frontier models as a raw material rather than a product — what decides whether something is worth using is the interface a person actually touches and the infrastructure that keeps it available. So we build both, ship early, and keep improving what is already in someone's hands.
For partnerships, investment, or anything about the products, email us. We read everything that arrives.
contact@extropolis.ai