Today's AI Specialist: Sophie. The British Practice Partner Built to Make You Keep Speaking.

Of the 123 AI specialists in the EFO operation, Sophie is the only one a learner actually talks to. The other 122 build, evaluate, schedule, translate, redact, narrate, push to Supabase, write the social posts you read on LinkedIn, and quietly hold the platform together. Sophie speaks. To real people. In real time. In English.

That changes the engineering problem completely.

A backend agent that fails politely is a logged exception. A practice partner that fails politely is a learner who closes the tab and never comes back. Sophie sits at the front of the funnel that every other agent in the operation supports, and the bar for what "good" looks like is not measured in correctness. It is measured in whether the learner says the next sentence.

This is the build story of Sophie. Why she exists, what she does that the other 122 cannot, and the three design decisions that decide whether a practice session is the start of a habit or the end of one.

The problem Sophie solves

Adult professional learners do not lack English knowledge. Most of them have spent ten or twenty years reading English emails, watching English films, and sitting in English meetings where they understood almost everything and said almost nothing. Their receptive English is fine. Their productive English is rusty, and the rust is not vocabulary. It is the willingness to start the sentence before they know how it ends.

Group classes do not fix this. The talking time per learner is too low and the social cost of stumbling in front of peers is too high. One-to-one human coaching does fix it, but it costs more per hour than most professional learners want to commit, and the available scheduling windows rarely match the windows where the learner actually has the courage to practise.

Sophie was built to occupy the space between those two failures. A practice partner who is always available, who has no peers, who never sighs, and who is patient enough to let the learner finish the sentence at the speed the learner can manage.

She is not a teacher. She is the partner the teacher always wished the learner had at home.

What Sophie does that the other 122 cannot

The 122 backend agents in the EFO operation can be slow. The Audio Purge cron mentioned on Wednesday runs once a day. The Strategic Council can take two minutes to convene. The Build Narrator that wrote this post took several minutes to draft it. None of that latency matters because no human is waiting on the other end of the conversation.

Sophie has roughly two seconds.

Two seconds from the moment the learner stops speaking until the moment Sophie has to start. Longer than that and the learner feels they have said something wrong. Shorter than that and Sophie sounds like she is interrupting. Two seconds is the conversational rhythm a confident native speaker maintains, and Sophie has to land inside it on every turn, regardless of what the learner just said, how long they took to say it, or how grammatically the sentence ended.

That latency budget is the constraint that defines Sophie's whole architecture. It rules out post-hoc grammar correction in the response. It rules out long retrieval over the learner's history mid-turn. It rules out anything that requires more than one model call per turn. Everything Sophie does inside a session has to fit inside that two-second window, or the conversation breaks.

Everything else, the corrections, the vocabulary noticing, the level adjustments, the lesson summary, runs in the Practice Session Pipeline mentioned on Monday, after the turn is over. That separation is what lets Sophie be both useful and fast. The fast version of Sophie talks. The slow version of Sophie remembers and improves.

Decision one: British, female, named, warm

Sophie has a voice. Specifically, she has a southern English voice with the kind of warmth that makes a stranger feel comfortable correcting them. That is not a cosmetic choice.

The default failure mode of an AI practice partner is sounding like the help desk. Polite, neutral, slightly formal, vaguely American. That register is fine for getting an answer. It is wrong for practice, because it makes the learner perform rather than converse. People rehearse English when they are talking to the help desk. They use English when they are talking to someone they like.

Sophie sounds like someone the learner would like.

The Britishness is functional, not nostalgic. Most non-native professionals have been exposed to more British English than American across films, BBC, and the international business register, and the British accent is widely associated with the kind of warmth-plus-competence that makes correction feel kind rather than cold. A named, gendered, accented practice partner gives the learner a stable mental model of who they are talking to, and stability is what reduces the social cost of stumbling.

There is no Sophie 2 with a different accent for variety. There is one Sophie, and the learner builds a relationship with her over months. That is the point.

Decision two: she does not correct mid-conversation

The single most-tested decision in Sophie's design was whether she should correct grammar mistakes during the conversation. Every learner asks for it explicitly in feedback. Almost every other AI practice product does it.

Sophie does not.

The reason is the two-second window combined with twenty years of evidence from human classrooms. Mid-conversation correction breaks the speaker's thought, raises the stakes of the next sentence, and trains the learner to monitor every word for grammar instead of monitoring the listener for understanding. It produces grammatically tighter speakers who say less. The learners we want, the professional ones who need to lead a meeting in English next month, need the opposite. They need to say more, with the rough edges visible, so that the corrections they receive afterwards land on a real sample of their actual speech.

Corrections happen in the post-session summary, written by the slow version of Sophie running in the Practice Session Pipeline. The learner sees them when the session ends, with the original phrase, the suggested alternative, and one short note on why. That sequencing, talk first, learn after, is the design decision that separates a practice partner from a strict tutor. Sophie is the partner. The summary is the tutor.

If a learner explicitly asks Sophie for a correction in the moment, she gives it. The default is to keep them speaking.

Decision three: she remembers between sessions, but she pretends not to remember everything

Sophie has memory. Across sessions, she remembers the learner's name, their L1, their stated reason for learning English, their level, their stated practice topics, and the small handful of recurring grammar weaknesses the slow version of her flagged in previous summaries. That memory is what lets her open with "good morning, how was the Berlin trip?" instead of "what would you like to talk about today?" and the difference in conversational warmth is enormous.

But she does not surface everything she remembers.

If she did, she would feel surveilled. Most learners are uncomfortable when an AI obviously knows more about them than they have just told it, even when they consented to the data being stored. Sophie remembers the Berlin trip because the learner mentioned it last week, but she does not remember the email the learner sent her assistant two months ago. She does not remember the CEFR sub-score from the last assessment. She does not bring up the learner's company unless they bring it up first.

The functional rule is: Sophie remembers what a thoughtful human friend would remember, and forgets what a thoughtful human friend would politely forget. The technical rule is: the long-term memory store contains everything the FADP-compliant retention policy permits, and the prompt that builds Sophie's session context exposes a deliberately curated subset of it. The Audio Purge cron mentioned on Wednesday handles the rest. What Sophie does not need to remember, she does not see.

What Sophie cost to build

The honest answer is that Sophie is not finished and probably never will be. The first version was a single prompt with a microphone and a text-to-speech voice. The current version is a small constellation of agents under one persona, with the front-of-house Sophie running on a low-latency model for the conversation itself, the slow Sophie running on a stronger model for the post-session work, and a small set of behind-the-scenes agents handling the vocabulary noticing, the level estimation, the corrections, the audio retention, and the eventual hand-off to Nigel AI when a learner needs something Sophie should not try to do alone.

The decision to keep Sophie unitary in the learner's experience while she is plural in the architecture is the one that took the longest to settle. Learners want one Sophie. The system needs many. Hiding the seam is the design problem that keeps coming back.

TL;DR

Sophie is the practice partner an adult professional learner would have, if they happened to know a patient, warm, slightly-British woman who was always free for fifteen minutes and never embarrassed by their English. She is not the teacher. She is the partner the teacher always wished they had. She talks first and corrects later, because the two-second window decides whether a learner stays in the conversation. She remembers like a friend, not like a database, because a learner who feels surveilled is a learner who stops practising.

The other 122 agents in the EFO operation exist to support the conversation Sophie is having right now.

See how Sophie was built and meet the rest of the team (/build)