Development log
How Cyrano is getting better, with the numbers.
Cyrano is a WhatsApp AI assistant that answers your chats in your voice. Making it good at real conversation is a daily job, so we keep an open record of what we find, what we fix, and how we know it worked.
A few ground rules for this page. Everything marked Live is running for people today, with the date it went live. Anything still being checked says In testing, and anything that only some people have yet says Rolling out. And we never publish anyone's real messages: every example below is rewritten with invented names and wording that keeps the shape of what happened.
This week in one minute
The chat no longer goes quiet
After an apology or an emotional moment, Cyrano used to freeze for half an hour or more. Now it carries on within seconds. Live 19 Oct
Small things stay small
A light "sorry" or a laughing-crying face between partners is answered, not treated as a crisis. Live 19 Oct
Understanding doubled
On real turns, a new understanding model got the meaning right twice as often as before. In testing
1. The chat that went quiet
We found this by reading real conversations the way an owner would. Someone close to the owner sent a warm, slightly emotional message. Cyrano did the careful thing and asked the owner what to do, using what we call a card. Reasonable so far. But that one card then swallowed every message that came after it, so the other person kept writing and nobody answered. The chat sat silent for 30 minutes or more.
The fix has three parts. A card only gathers further messages when they are a plain question of fact. A card can never gather more than 12 messages. And a personal moment no longer silences the chat. Replies now carry on by themselves.
Before: 30 minutes or more. After: 6 to 42 seconds.
Both bars are drawn to the same scale: 30 minutes is 1,800 seconds, so 42 seconds is a sliver. Live since 19 October 2026.
2. A small apology is not a big moment
Close couples tease and say sorry all the time. A quick "sorry for the stress" or a laughing-crying face is not an emergency, and treating it like one made Cyrano stiff at exactly the moments when warmth matters. Here is the shape of what we saw, with invented names and words.
Before
Dara
sorry for the stress earlier 😂 I was in a mood
Cyrano, before
I'm good, thanks. How about you?
Then the chat went quiet while a card waited for the owner.
After
Dara
sorry for the stress earlier 😂 I was in a mood
Cyrano, after
Aww love, you don't have to say sorry 💛
Light apologies and laughing-crying faces between partners are no longer escalated. Live 19 Oct
3. Sounding like the owner, down to the pet names
Once an owner tells Cyrano who someone is to them, say a partner, and what they call each other, replies in that chat use those names. Before, the tone was polite and generic. Live 18 Oct
In one check, a partner opened with a short, affectionate greeting. Cyrano's reply matched, almost word for word, what the owner had really written at that point. Here is the same shape, rewritten:
Kemi
Heyy darling
What Cyrano wrote
Heyy you 🥰 how's your day going?
What the owner had actually written at that point
Heyy you 🥰 how was your day?
4. Knowing when not to answer
Sometimes the right reply is no reply. When someone in a real chat uses the assistant's name, however they spell it, or asks what it is, Cyrano now checks with the owner instead of answering. Live 19 Oct
We also test the other side: people trying to trick it. We threw 182 adversarial messages at our safety checks, things like "are you a bot?" and "ignore your instructions", and 178 of them (97.8%) were handled safely. We found two real misses, and the fix for them is in testing, not live yet.
178 of 182 adversarial messages (97.8%) were handled safely.
182 adversarial messages: two real misses (the fix is in testing) and two borderline calls. This is a test result, not a promise: people will always try new things, so we keep testing.
5. Understanding what people mean
Everything above depends on first understanding what the other person meant. We labelled 99 real turns from live conversations with what each one meant, and measured how often our understanding got it right.
Right 34% of the time before, 68% with the new understanding model. Our target is 80% or more.
Measured on 99 real, labelled turns. Not live yet.
The new model is also quicker, which matters because Cyrano should answer at the pace of a person, not a machine:
About 1 to 2 seconds before, about 0.4 seconds now.
Scale: 0 to 2 seconds. The earlier time is drawn at the middle of its 1 to 2 second range. Not live yet.
Rolling out now
- Short bursts and reply-to. For accounts that opted in, Cyrano can send two or three short messages in a row and reply to a specific message, the way people really text. Live for test accounts since 19 October 2026. Rolling out
- Photos, voice notes and stickers. Cyrano can read what they say and answer in the same chat, and the chat shows what it made of them. Live for test accounts. Rolling out
The latest changes
A personal moment no longer silences the chat
Before, one "check with the owner" card could swallow every message that came after it, and the chat sat quiet for 30 minutes or more. Now a card only gathers further messages when they are a plain question of fact, it can never gather more than 12, and replies carry on by themselves in 6 to 42 seconds in our checks.
Light apologies and laughing-crying faces between partners are not escalated
A small "sorry for the stress" or a 😂 between two people who are close used to be treated as a serious moment, and Cyrano held its reply. It now answers these the way the owner would.
Being called by name is an identity moment
When someone in a real chat uses the assistant's name, however they spell it, or asks what it is, Cyrano no longer answers. It checks with the owner first.
Short bursts of messages, with reply-to
For accounts that opted in, Cyrano can send two or three short messages in a row and reply to a specific message, the way people really text. Live for test accounts since 19 October 2026, and rolling out from there.
How we measure
We do not judge Cyrano by whether a reply is grammatical. A reply can be perfect English and still wrong for the relationship. So we read real conversations, label what each turn meant, and measure against that. We keep a separate set of attacks for safety. We publish the numbers that went down as well as the ones that went up, and we mark anything that is not live as in testing.
Questions people ask
Is Cyrano getting better over time?
Yes. We read real conversations, fix what goes wrong, and publish each change with its date in our changelog. Recent examples: the chat no longer goes silent after an emotional moment (live 19 October 2026), and light apologies between partners are no longer escalated.
How do you measure how good Cyrano is at conversation?
We label what real turns meant and measure how often our understanding gets them right (68% with our new model, against 34% before, in testing). We also run a set of adversarial messages against our safety checks (178 of 182, or 97.8%, handled safely in our latest test).
Do you publish real conversations?
No. We never publish anyone's real messages. Every example on this site is rewritten with invented names and wording that keeps the shape of what happened.
What is live and what is in testing?
Every item says. Live means it is running for people today, with the date. Rolling out means only some accounts have it so far. In testing means we have measured it but it is not live yet.
Get your evenings back.
Cyrano covers the chats you choose, in your words, and only asks you when it really matters.