An endless, AI-generated conversation between Werner Herzog and Slavoj Žižek — released before ChatGPT existed. Built to show that synthetic voice could be uncanny, funny and genuinely worth listening to, and to make an argument about deepfakes by example rather than assertion.
(a voice, alone, in a room full of servers)
Machines are learning to speak. Most still sound like machines.
Silicon Soliloquy is a studio for voice interfaces and synthetic personas — systems people talk to, and then keep talking to. I have been building convincing machine voices since 2022, back when it took a year of work to get a sentence that held up.
What I work on
Voice interfaces
Real-time systems that listen, interpret and answer. Turn-taking, latency and graceful repair decide whether people keep speaking — far more than the choice of model does.
Synthetic personas
Characters with a consistent voice, a memory and a point of view. Convincing enough to survive the first ten minutes, when the novelty has worn off and the illusion usually breaks.
Problems without a method
Briefs where nobody is yet certain the thing can be done. I build the prototype that answers the question, early, before the budget is committed to an approach.
Selected work
A formula language, generated
2025Client work for a survey platform: turning plain descriptions of questionnaire logic — show this question if they picked two or more departments — into formulas in the platform's own language, a DSL of sixty-odd functions with a real type system. The generation was the easy half. The work was proving the output correct, with a type checker, an interpreter and generated test data, because a formula that merely looks plausible is more dangerous than none at all. The opposite problem from a persona: not convincing, but right.
Aleph
in progressA literary companion in your pocket. Co-founded, and currently the thing that occupies most of my week.
Working together
Most engagements begin small: a short discovery phase to work out whether the idea holds together and where the genuinely hard part is hiding. From there the shape of the work becomes obvious.
I lead the work myself, and bring in collaborators I have built things with before when the scope asks for more hands. Work has ranged from a prototype that answers one question to systems a company's product depends on.