When a person is no longer in this world, their loved ones always miss the way they spoke, thought, and joked.
Novessa is building a way to keep that from being lost: a system modelled on how one specific person spoke and thought, so the people who loved them can still hear how they'd have answered. Not a copy, not bringing anyone back. Built only with the family's consent, and always labelled as what it is.
The Black Mirror comparison deserves a serious answer.
People compare this to Black Mirror, and the concern underneath the comparison is a reasonable one: that a system standing in for someone who has died could deepen grief rather than ease it, or slowly take the place of the person in the minds of those who loved them. That is a real risk, and we treat it as one.
The story that raised it turned on a system that presented itself as the person, with no one able to switch it off. Novessa is built the other way around. A presence is always labelled as what it is, it is created only with the immediate family's consent, and any family member can have it permanently deleted. It is not meant to stand in for the person or to be mistaken for them.
The research question is also narrower than resurrection. Can the particular way one person thinks be modelled well enough to be recognisable? How they reason. What they find funny. What they notice, and what they let pass. We do not know yet. That is why it is research and not a product.
Two write-ups, both early.
We write up the parts of the work that are far enough along to state clearly, even when the answer is still open. Neither is peer-reviewed or public yet. We will link each one here once it is.
Voice as a detection channel
Whether a synthesised voice carries a signal, in the audio itself, that it is synthetic. Currently being written and prepared for submission. Not finished, not public yet.
Measuring recognition fidelity
How you score whether a modelled presence is recognisable to the people who knew the person: what to ask them, how to run it blind, what counts as a pass. Early-stage: a problem statement and a first method, no results yet.
The applied work funds the research.
Novessa is two things, run by the same team. We build agentic AI systems and real-time voice infrastructure for companies. That is commercial work with commercial customers, and it pays for the research.
We say this plainly because a research company with no revenue is a hobby, and a consultancy with a moonshot bolted on is a distraction. The order matters. The applied practice exists to fund the question.
Their voice first. Presence next. A physical body, much later.
Voice
Reconstructing the voice of the person who died: how they sounded, their cadence, the way they landed a joke. Not a generic product voice, the specific voice of one person who is gone. This is the part of the work that is furthest along.
Presence
A personality model stable enough to hold a long conversation, and recognisable to the people who actually knew them. Open problem. We publish as we go.
Embodiment
Embodiment means a physical, robotic body, something you could be in a room with rather than only text on a screen. We have not built this, we are not implying we have, and it is years away. It is where the research points, not something you can buy.
Be the first to know.
Leave your email and we will tell you the day the prototype is ready to try, plus anything else worth telling you before then. No spam, no countdown.
By joining you agree to our Terms and Privacy Policy.
Technical roles, every one paid.
Every role is paid.
Applied research across agentic systems and evaluation. You will move between commercial builds and the research problem behind them, depending on the week.
Turning a working prototype into something that survives real traffic. Model selection, RAG pipeline tuning, and the unglamorous work of making outputs consistent.
Low-latency speech pipelines for commercial clients. Barge-in, turn detection, telephony integration, and the parts of a voice stack nobody notices until they break.
The core research question, shared with the research side of Novessa. Building a stable behavioural model out of source material that is sparse, messy and personal.
Total ownership of design, across both the research presences and the commercial work. Everything we make gets touched by someone relying on it. If that makes you more careful rather than less, we should talk.