Behind the Build
Field notes from the inside of an AI-native operating company. One founder, four divisions, 123 AI Specialists. Wins, mistakes, and the architecture decisions in between — dated and real.
The Growth Advisor's lens is one question. Where in your funnel is revenue leaking, and what specific AI mechanism captures it? She does not set the price. She does not approve the spend. She names the leak, names the mechanism, and gives Margin the range to test against.
Read the post →After the five advisors have written their lenses, Peer Review strips the names, shuffles the labels, and runs the anonymous round where each advisor has to critique four others without knowing whose work they are reading. The whole council depends on her not telling. Here is the seat that earns the council's fee.
Read the post →The Collective Brain has 412 files. I read maybe 40 of them per week. That gap is dangerous. So I built an agent whose only job is to monitor the Brain and tell me, every Monday, what I have missed that matters.
Read the post →The Margin Advisor reconstructs the math. The Growth Advisor names the mechanism. The Capacity Advisor asks the question both of them are not allowed to ask: can the team actually run this without breaking? Here is the seat that gives every Top Priority a Green, Amber, or Red flag and a named owner.
Read the post →Four advisors kept producing plans that were intellectually right and operationally undeliverable. So I built a sixth Council seat whose only job is to ask 'by whom, when, against what.' Here is how The Operator works.
Read the post →The Strategic Council runs in two surfaces. The founder-facing one is chaired by the Chairman, who closes a process already disciplined by Red Team, Peer Review and schema validation. Here is how the seat actually works.
Read the post →Every recommendation that reaches the Chairman has been through the Margin Advisor first. He reconstructs the math, runs TCO across twelve months, and classifies the work as Quick Win, Strategic Bet, or Reject. Here is the seat whose job is to be unimpressed by the revenue number until the cost side has been priced.
Read the post →Every practice session on the platform ends with a feedback summary. The Spoken Feedback Agent writes that summary — not in the voice of a teacher grading a test, but in the voice of a coach noticing what the learner did well and where they can practise next. The tone correction was the hardest part of the build.
Read the post →The Strategic Council has five advisors. The Council Director decides which of them get to speak on a given question, in what order, and which never see the brief at all. He is the agent who filters out the advisors who would have noise to add, before they add it.
Read the post →When I built the fluency assessment, I had to make a single design decision that everything else flowed from: when fluency and accuracy disagree on a sample, which one wins? Most assessments score accuracy. We score fluency. Here's the story of why, and what it costs us.
Read the post →The Custom Conversation Pack we send every new consultation lead includes a personal video from me when the pack warrants it. The Loom Personalisation Agent is the agent that decides which packs warrant one, what scenes to include, and what to say — without ever recording a frame itself. The agent owns the brief, not the camera.
Read the post →Every advisor in the Strategic Council I built last week has one thing in common: they are paid, in effect, to give me good news. The Red Team is the only agent in the Council whose only job is to argue against me. It is the agent I built specifically because I had stopped being able to tell, by myself, whether my own plans were good.
Read the post →A financial operating intelligence layer for a specific founder decision: where the next franc has to go. Six agents, triple-auth wall, Swiss SME context.
Read the post →How a single editor agent gates an AI content operation: owns the calendar, dispatches briefs, reviews every draft against a short voice standard.
Read the post →Of the 123 AI specialists in the EFO operation, Sophie is the only one a learner actually talks to. A practice partner that fails politely is a learner who closes the tab and never comes back. 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.
Read the post →Sophie is the practice partner. Nigel AI is the expert coach clone for the conversation Sophie cannot have. The architectural decision the platform would not work without.
Read the post →Every piece of paid acquisition on EnglishFluency.Online runs against an ICP defined by one agent. The Audience Architect is the agent who decides who the platform should be marketing to — and, more importantly, who it should not. The decision shapes every ad, every landing page, every dollar of spend downstream.
Read the post →When a new learner arrives on the platform, the Onboarding Placement Agent runs the first session — the one that decides where the learner gets dropped into the programme. Wrong placement means months of wasted practice. The agent has three minutes of audio and the rest of the learner's intake answers to get it right.
Read the post →Every learner on the platform records a 60-second speaking sample at regular intervals. The Snapshot Fluency Analyst is the agent that scores it — four sub-scores, in roughly twenty seconds, with the level of precision that a human assessor would take twenty minutes to produce. The design problem was not the scoring. It was the trust.
Read the post →Sophie's voice is what the learner actually hears. Coach Speak is the agent that turns Sophie's text into audio in under a second, while preserving the warmth, the British accent, and the slight imperfection that makes her sound like a person rather than a system. The latency is the build problem. The imperfection is the design problem.
Read the post →A real advisory board costs CHF 8-15k and weeks. The Strategic Council returns the same in twelve minutes, with structure designed against convergence.
Read the post →I deleted our lead qualification form. The Diagnostic Chat Designer replaced it with a bilingual AI conversation that qualifies and demos at the same time.
Read the post →The Strategic Council surfaced a Swiss FADP audio-retention exposure I was not tracking. Within days, a nightly data-protection agent closed the gap.
Read the post →Sophie is the practice partner the learner talks to. But Sophie does not pick the conversation topic — she pulls it from the Conversation Partner agent. The Conversation Partner is the agent who decides, for each learner, what conversation they should have next, and the design decisions behind that choice are why Sophie's sessions feel personal.
Read the post →Every post on EnglishFluency.Online ships with translations into Italian, French, and German, a vocabulary list with CEFR levels, three grammar notes, and five comprehension questions. The agent who produces all of that is the Language Analyst — the one who turns each blog post into a piece of learning material, not just a piece of marketing.
Read the post →Every post on EnglishFluency.Online has a writer who decided what it says, an editor who decided whether it ships, and an SEO Specialist who decided whether anyone will find it. The third decision is the one most content operations get wrong, because they treat SEO as decoration on the writing. The SEO Specialist treats it as a parallel design problem.
Read the post →Nine specialist agents built EFO's eduQua:2021 quality dossier in three weeks. The traditional consulting cost was CHF 25,000-60,000. The architecture is also now the live QMS.
Read the post →Every booked consultation completes a six-exchange diagnostic chat before the call. The conversation feeds a personalised Conversation Pack and gives me three minutes of context I would not otherwise have. Here is the agent who decides what those six exchanges have to do.
Read the post →Every reasonable instinct said one high-context LLM call should beat seven specialist agents in sequence. The instinct was wrong, for reasons that compounded the moment the system was producing 70 posts a week. Here is what made me reject the one-shot model.
Read the post →Most language courses are built by sequencing what is convenient to teach. The Course Strategist refuses that. She sequences from what the learner has to be able to say at the end, back through the scenarios that get them there, and only then asks what each lesson should contain. Here is the architect of every EFO programme.
Read the post →I used AI for two years before I noticed that I was having AI conversations and calling them decisions. The decisions did not survive the next morning's coffee. So I rebuilt the surface as actual meetings. Here is what changed, and the day I knew the change had worked.
Read the post →I held an eduQua certification for a training school in Lugano for one renewal cycle. The auditor said one sentence that has shaped every quality system I have built since. Here is the moment, and the dashboard that exists because of it.
Read the post →Between the moment a prospect clicks the Book Call button and the moment they show up on the call, twenty-one things happen. For three months nobody owned the sequence. So I built an agent whose only job is the choreography of every step in the journey.
Read the post →Why I built a seven-person team whose only job is to keep six living maps of how the company actually reasons, decides, and learns. The Cartography Team Leader is the one who keeps those maps in sync.
Read the post →Every blog post in the build needs an editorial portrait that feels like the same hand made it. I tried to do this with prompts. Then I tried to do it with humans. Then I built an agent whose only job is to hold the visual identity together. Here is how the Image Library Steward works.
Read the post →
