The Lab · Conversational AI
Meet Eva. An AI you can actually talk to.
Eva holds a real-time voice conversation the way a person does. She answers in under a second, lets you interrupt, remembers what you said earlier, and keeps the conversation on the subject she was given — from the first minute to the last.
In plain words
What talking to Eva feels like.
Most voice assistants are good at one question and one answer. A conversation is something else: it has a rhythm, interruptions, tangents and a memory of its own. Eva is built for the conversation.
She keeps up
Her first words arrive in under a second. She waits until you have actually finished, and a quick “mm-hm” from you means “keep going”, not “stop”.
She remembers
She remembers what was said earlier in the conversation — and only what you actually heard. Cut her off mid-sentence and she knows you never heard the end of it.
She stays on subject
Give her a topic and a few goals. She will follow a tangent the way a person would, then bring the conversation back to where it needs to go.
She knows what she doesn’t know
When she needs a fact, she checks it. When she can’t find it, she says so instead of making something up.
The problem
Short answers are easy. Long conversations are not.
Ask a voice assistant one question and it usually does fine. Keep talking for half an hour and the cracks show:
- Silence. A two-second gap feels broken on a call, yet most systems wait for a complete answer before saying a word.
- Forgetting. Details from ten minutes ago slip away — or worse, get contradicted.
- Drifting. Without a sense of where the conversation is going, it wanders off or circles back to the same point.
- Talking over you. People interrupt, agree out loud and pause to think. A system that can’t tell those apart either barges in or never stops.
Eva treats a conversation as something live and shared, not as a series of separate requests. That one idea shapes everything below.
How Eva is built
Two halves: Eva-talk and Eva-brain.
Eva is split into two parts that do very different jobs.
Eva-talk is the part in the room. It listens, decides when it is her turn, thinks on its feet and speaks. Its job is to be fast and natural.
Eva-brain is her long-term memory — a careful library. It knows what is true, since when, and who said so. Its job is to be right.
Keeping them apart means the voice never has to slow down to be careful, and the memory never has to cut corners to be fast.
Eva-talk — in the room
Eva-brain — the library
- You
- The person Eva is talking to — or several people at once.
- Speakers are recognised by voice, so in a group a remark addressed to someone else by name is not mistaken for an interruption.
- Ears
- Turn your voice into text while you are still speaking, word by word.
- Streaming transcription with its own end-of-turn signal. A separate listener reads tone of voice, and only influences Eva when it is confident.
- Turn-taking
- Works out whether you have finished, paused to think, agreed, or cut in.
- Three separate decisions: end of turn, how long to pause before answering, and agreement versus interruption. Silence is judged against each speaker’s own pace, and a quiet person is told apart from a dropped line.
- Director
- The conductor. It starts Eva’s turn and keeps everyone else in step.
- Builds each turn from the conversation so far, the brief, Eva’s profile and background notes from Eva-brain, then starts the quick thinker and the side channels in parallel.
- Quick thinker
- Writes what Eva says, in her own voice, as fast as she can say it.
- A fast model whose only job is talking. It never blocks: when it needs a fact it asks the researcher mid-sentence and carries on with a line that commits to nothing until the answer arrives.
- Speech queue
- Holds the next few phrases, ready to play — or to cancel.
- Each piece moves from drafted, to ready, to playing, to spoken. Anything not yet playing can still be cancelled or reordered, which is what keeps interruptions and corrections clean.
- Voice
- Speaks, with an expression that fits the moment.
- Speech is produced in short chunks and streamed, so the first phrase plays while the rest is still being written. The range of emotion is bounded by the tone of the conversation.
- Conversation memory
- Remembers the conversation — but only the words that actually reached you.
- A transcript of what was played, not what was planned, plus a running summary, the facts established so far, the threads still open and the agenda.
- Overseer
- Steps back and thinks about where the conversation is going.
- A slower model that runs behind the conversation. It keeps the running summary, gathers verified facts, anticipates likely questions and suggests a direction for the next turn — a direction, never a script.
- Auditor
- Quietly reviews each reply and flags anything that needs fixing.
- Deliberately a different model family from the quick thinker, so the two do not share blind spots. It also names the current subject and tracks the tone. At most one correction per turn, so Eva never spirals.
- Researcher
- Answers the quick thinker’s questions, using Eva-brain.
- Starts the moment a question is asked, while Eva is still talking. “Nothing found”, “nothing on your account” and “can’t confirm” are kept as three different answers.
- Secure bridge
- The only door between the two halves.
- An authenticated API. Who the customer is comes from the call itself, never from the model, and access fails closed: when in doubt, the answer is no.
- Sources
- Everything Eva is allowed to know: documents, product data, her own background.
- Documents are read in any language. How far a source can move a belief depends on how much that source is trusted.
- Claims
- Every statement, kept exactly as its source made it.
- Claims are never edited, and each one points back to the exact passage it came from.
- Facts
- What Eva currently believes, weighed from the claims.
- A fact can be replaced when new evidence arrives, but never deleted. Each one records when it was true, when Eva learned it, and when the source said it.
- Search, as of now
- Finds the answer and shows where it came from.
- Answers reflect what is true today, cited down to the passage. When the evidence does not settle a question, it says so instead of picking a side.
- Profile and opinions
- Eva’s background and opinions, so she is the same Eva in every conversation.
- Loaded when a conversation starts, so her personality never has to be looked up mid-sentence.
Anatomy of a turn
A few seconds of conversation, slowed down.
Everything below happens in a few seconds of real conversation. Pick a situation and watch who does what, and when. Notice that Eva starts speaking long before the slower parts have finished their work.
What Eva remembers
A simple reply
A podcast. You are the host; Eva is your guest.
- 0.0s · You You ask a question. Eva hears it as you say it, not after.
- 0.2s · Ears Your words become text while you are still talking, so nothing has to wait for the end of your sentence.
- 2.1s · Ears Turn-taking decides you have actually finished — not just paused to think.
- 2.2s · Quick thinker A short, human pause before answering: none for a greeting, a little longer for something serious.
- 2.6s · Quick thinker Her first words arrive in under a second. She speaks phrase by phrase while the rest of the answer is still being written.
- 3.0s · Auditor In the background, an independent model reviews what she is saying: is it consistent, still on subject, in the right tone?
- 3.9s · Memory & agenda As each phrase finishes playing, it is added to memory — what you heard, not what was planned.
- 5.2s · Memory & agenda The second phrase lands in memory the moment it has been played.
- 5.3s · Auditor Nothing to fix, so nothing happens. The check never slowed the reply down.
She checks a fact
A support call. You are a customer; Eva is on the help desk.
- 0.0s · You You ask about your own account.
- 0.2s · Ears Transcribed as you speak.
- 1.7s · Ears You are done; it is Eva’s turn.
- 2.1s · Quick thinker She starts talking straight away. This holding line can react, but it is not allowed to state a date, a name or a number she has not been given.
- 2.3s · Researcher Mid-sentence, the quick thinker hands the question to the researcher. The lookup starts before her first phrase has finished playing.
- 2.5s · Eva-brain Eva-brain checks who is calling — from the call itself, never from the model — and searches only what that customer is allowed to see.
- 3.5s · Researcher The answer comes back with its source. Had nothing been found, Eva would say she can’t confirm it, not guess.
- 3.7s · Quick thinker The rest of the holding line is dropped, and Eva continues from exactly what she already said, so the sentence joins up naturally.
- 4.0s · Auditor The auditor reviews the answer in the background.
- 5.9s · Memory & agenda The exchange is saved as heard, and the renewal date becomes an established fact for the rest of the call.
You interrupt
The same podcast. Eva is in the middle of an answer.
- 0.0s · Quick thinker Eva is explaining a point.
- 0.9s · You You say “mm-hm”.
- 1.1s · Ears From the first sound, Eva decides this is agreement, not an interruption — so she keeps going.
- 2.6s · You This time you really do cut in.
- 2.8s · Ears The first words say “I want the floor”. Eva gets ready to yield.
- 3.4s · Quick thinker She doesn’t stop mid-word like a machine: she finishes her clause, then yields. The rest of what she planned is cancelled before it is ever played. A hard stop is kept for when you say “stop”.
- 3.5s · Memory & agenda Memory keeps only the words that reached you. As far as the conversation is concerned, the rest never happened.
- 4.7s · Ears You have finished your question.
- 5.1s · Quick thinker She answers the question you actually asked, not the one she was busy with.
- 7.2s · Memory & agenda And the answer is saved, exactly as heard.
The conversation drifts
An interview. Eva is the interviewer, with goals to cover: your first customer, your biggest mistake, what comes next.
- 0.0s · You You drift into a story about your dog.
- 2.7s · Ears You have finished your thought.
- 2.8s · Memory & agenda The agenda is run by plain code, not by the model. The current goal — your first customer — is not covered yet, so the move is: acknowledge first, then bridge back.
- 3.0s · Auditor The auditor notes that the subject has moved, and names the new one in a single line.
- 3.2s · Quick thinker Eva follows you, warmly and briefly — then uses the tangent as the way back to the goal.
- 4.8s · Overseer One turn behind, the overseer folds the story into the running summary and suggests what should follow this goal — a direction, never a script.
- 7.8s · Memory & agenda Her reply is saved as heard.
- 8.0s · You You are back on track — and it didn’t feel like being steered.
- 10.1s · Memory & agenda Once you have answered, the goal is marked as covered and the agenda moves on to the next one.
Under the hood
For the technically curious.
Here is how the pieces fit together: the big picture, without the recipe.
Talk now, think in parallel
A person doesn’t stop talking every time they need to remember something. Neither does Eva. Each part of a turn is handled by a model with a single job, and they run at the same time:
- a quick thinker that only talks, fast enough to start in under a second;
- a researcher it can call on mid-sentence when it needs a fact;
- an auditor that reviews each reply after it is spoken and can trigger a correction;
- an overseer that runs one turn behind, keeps the summary and suggests where to go next.
When the quick thinker asks for a fact, it keeps talking — but with a holding line that is not allowed to commit to a name, a number or a verdict. When the answer arrives, the unspoken rest of the holding line is dropped and Eva picks up from exactly what she has already said.
Turn-taking is three decisions, not one
“When should I talk?” sounds like one question. Eva treats it as three. Has the speaker finished, or only paused? How long should I wait before answering? And when they speak while I am talking — is that agreement, or an interruption?
She talks through a “mm-hm”, yields to a real interruption by finishing her clause rather than cutting off mid-word, and judges silence against each speaker’s own pace. In a group, a remark addressed to someone else is not treated as an interruption. She can also tell a quiet person from a dropped line.
Memory is what was heard
Most systems remember what they generated. Eva remembers what was played. Her speech is queued in short pieces, and each one is written to memory only once it has actually reached your ears. If you interrupt, memory keeps the words that got out; the rest never happened.
To stay sharp over a long conversation, older turns are folded into a running summary, next to a short list of the facts established so far and the threads still open. The conversation can run long while what Eva carries into each turn stays small.
Staying on subject
Every conversation starts with a brief: the premise, who is in the room, the goals — some marked as must-cover — the topics to steer around, and hard lines that are enforced in code rather than left to a model’s judgement.
As the conversation moves, the auditor names the current subject in a single line and keeps track of the tone. The agenda itself is run by plain code, one goal at a time: Eva acknowledges what you just shared, moves on only once a goal is covered, and wraps up only when every must-cover goal is done.
Models judge; code decides. It is the same principle behind the Control Model.
Eva-brain: a memory that knows when
Eva-brain keeps two kinds of knowledge apart. Claims are what a source said: they are never edited, and each one cites the exact passage it came from. Facts are what Eva currently believes: they can be replaced when new evidence arrives, but never deleted.
Every fact tracks three kinds of time: when it was true, when Eva learned it, and when the source said it. That is how Eva answers “as of now” — and how she can explain what changed. When the evidence does not settle a question, Eva-brain says so instead of picking a side.
Access is decided in one place only. The person Eva is talking to is identified by the call itself, never by the model, and anything uncertain is refused by default.
Many models, no single point of failure
Eva does not depend on one model. Each role uses a model suited to that job — the auditor is deliberately from a different family than the quick thinker, so the two do not share blind spots — and the roles can fall back to other providers to keep the conversation flowing.
Every side channel is also built to fail gracefully. A late correction is dropped. A late note waits for the next turn. A slow lookup means Eva answers carefully on her own and corrects herself afterwards if she needs to. The conversation never waits for its slowest part.
Measured, not guessed
Conversation quality is easy to argue about and hard to measure, so we measure it. Scripted conversations — interruptions, tangents, long silences, tricky questions — replay on a simulated clock, quickly and identically every time. Every real session produces a timeline and metrics such as time to first word, dead air and interruptions handled, and every model call is logged.
In use
One Eva, many rooms.
The same two halves adapt to very different conversations. The brief changes; the discipline doesn’t.
Podcast guest or host
A guest with opinions of her own, or a host who listens, follows up and keeps the episode on track.
Interviewer
Works through a set of goals within a time budget, following up on what you actually said.
Support and sales desk
Answers from your product and account data, and says “I can’t confirm that” rather than guess about your account.
The voice of a book
Speaks for an author or a story, grounded in the text itself.