2 Mar~1,200 hrs
The EFO platform
The app learners log into. Ten runtime agents serve traffic — voice lines, fluency scores, community spotlights — each with its own credit logic, rate limits and refunds.
3 Mar~600 hrs
The course-production pipeline
Seven agents turn one topic, CEFR level and pathway into a complete pedagogical package: structure, lesson grids, scripts, dialogues, assessments and QA.
3 Mar~400 hrs
The coaching + personalisation pipeline
Five agents turn raw learner activity into a personalised plan and a coaching narrative: profile, personalise, then review progress.
3 Mar~120 hrs
The cross-cutting copywriters
Two general-purpose marketing agents any other project can call when it needs copy in the EFO brand voice.
3 Mar~120 hrs
The community + support pair
Two agents: an inward briefing on what is happening in the community, and outward spotlights and replies. Both in Nigel's voice, neither sends without his review.
Mar~200 hrs
The in-app practice surfaces
Three more production agents inside the app — practice, drill and push — serving learners every day.
Mar~60 hrs
The email-delivery runtime
The runtime that pairs AI-drafted email content with Resend delivery.
Mar~150 hrs
The customer-touchpoint runtime
Three production agents handling the actual delivery of customer-facing support and welcome messages.
Mar~500 hrs
The assessment + placement engine
Six agents run in sequence when a candidate hits the assessment to produce a CEFR-aligned result. The gate every paying learner passes through.
Mar~300 hrs
The blog-rendering pipeline
Turns raw markdown into a fully shipped post: hero image, audio narration, floating bubble, OG image and learner-enrichment data.
Mar~180 hrs
The trial + demo runtime
The unauthenticated trial flow: own rate limits, self-contained, ending in a trial report. The surface every prospect touches before they assess.
17 Mar~800 hrs
The daily content engine
Eight agents shipping a blog post and four matched social posts a day, plus the Behind the Build stream — seven posts and twenty-eight socials a week, every week.
7 Apr~500 hrs
The Fluency Clinic
Six agent stacks running the premium Clinic in parallel: prep packs before, the recording flow during, the assets that ship after. Bilingual EN/IT throughout.
8 Apr~900 hrs
The eduQua dossier
A complete eduQua:2021 quality-management documentation suite, built in three weeks. A real institution would budget six to twelve months and a five-figure engagement.
12 Apr~600 hrs
Traffic recon & capture
Eleven agents in a strict SEO and AI-search pipeline. Adding a new target is a content-only task for the writers; the engineer only touches code for new features.
Apr~120 hrs
The branded transformation dashboard
One agent: a prospect submits a nine-question quiz and their URL; it fetches their site, applies their brand, and generates a personalised AI-transformation report.
Apr~300 hrs
The productised Council
The same eight Council agents, productised as a paid customer-facing assessment. We sell the council we use ourselves.
Apr~150 hrs
The live QMS dashboard
The live admin surfaces that turn the eduQua dossier into usable dashboards: corporate, documents, learners, quality and satisfaction.
Apr~250 hrs
The conversion runtime
Four production agents that run when a prospect hits the funnel — the implementation of the conversion spec.
Apr~150 hrs
The production tooling
Python build pipelines, not LLM agents, that turn agent-generated content into shippable PDFs and slides.
Apr~120 hrs
The Brain MCP server
Exposes the Collective Brain as a remote tool server over Tailscale, so any MCP client can consult it from anywhere — not only a session opened in the Brain folder.
19 Apr~100 hrs
The Brain Health agent
The Brain's immune system. Runs nightly at 23:00, audits pending updates, executes overnight writes and files a daily report — catching drift before it compounds.
19 Apr~700 hrs
The Vault
A Swiss recovery-grade financial-intelligence layer: three AI advisors plus a Creditor Claim Evaluator and subsystems. Decision support under financial constraint, not bookkeeping.
20 Apr~600 hrs
The accounting pipeline
A full double-entry pipeline scaffolded in one session: five source types into a routing engine, three-tier FX, validated entries against a 29-account chart. 2023 reconstructed in two days, zero unclassified.
26 Apr~300 hrs
The internal Strategic Council
The eight Council agents operating internally, used by Nigel on his own decisions across the ecosystem. This was the prototype; the MBA product was the lift.
26 Apr~200 hrs
The compliance layer
Three cron agents that turn the privacy policy into behaviour: audio-retention purge by article, lifecycle crons and client-error reporting. Retry-safe and secret-gated.
27 Apr~400 hrs
The conversion engine
Eight agents on five named levers — dashboard visibility, ad creative, diagnostic pre-qualification, show-up rate and dignified deflection — targeting consistent MASTERY enrolment month after month.
30 Apr~200 hrs
The Behind-the-Build feed
The public build feed and full /build post rendering: audio narration, three-language analysis and learning materials on every story.
5 May~700 hrs
Brain v2 financial reconstruction
2023 reconciled to the signed bilancio to the centime, and a full 2024 reconstruction from Stripe, PayPal, bank and payroll, with live P&L and balance-sheet pages.
11 May~400 hrs
The Conversation Pack engine
Personalised pack generation on Sophie's voice: key-phrase audio, a 'pack from notes' path, and re-takeable, ad-hoc diagnostics feeding each pack.
13 May~300 hrs
Brain security + financial isolation
MFA/AAL2 on the Brain, the financial brain moved to a dedicated allow-listed Supabase project, and full-funnel diagnostic visibility with a Loom review queue.
17 May~400 hrs
Clinic demos + auto blog analysis
Sophie clinic demos and a clinic library, automatic three-language blog analysis (vocabulary, grammar, comprehension), and assessment form-variant assignment.
19 May~500 hrs
The agent persona system
Portrait-first identity locking with Nano Banana Pro, locked agent-group ensembles and batch persona invention — the system that gives every specialist a consistent face.
20 May~400 hrs
The talking-head video pipeline
A HeyGen Avatar IV pipeline and team-card video studio that turn any specialist's portrait and script into a spoken introduction.
21 May~400 hrs
The operating-company showcase
The public /team and /agents pages: the live roster, the Strategic Council in their own words, today's specialist, and this build log.
22 May~150 hrs
The organisation cartography system
A seven-agent meta-team that maps the whole company to itself across six dimensions, registry, competency, consultation, accountability, work medium and customer journey, kept current as the org grows, with a value-stream overlay that groups every specialist by the outcome it serves.
24 May~160 hrs
The agent voice standard
A capability-not-authorship pitch standard and validation pipeline that gave all 143 specialists a public introduction in one consistent, accuracy-checked voice, generated graph-first from the cartography maps and verified against the Collective Brain.
26 May~60 hrs
The asset-production platform
The shared production engines registered as one platform team: a document render engine behind every PDF, an image library and alt-text pipeline, and the producer that turns each specialist's pitch into a team-page video.
27 May~90 hrs
Word-synced agent-video captions
Captions across the agent pitch-video library: Whisper word-level timestamps drive a WebVTT generator, a paginated backfill route caption-stamps the whole back catalogue, and the player reveals each spoken word in a rolling karaoke window. An accessibility and comprehension layer on every specialist's introduction.
28 May~120 hrs
The founder walkthrough
An interactive walkthrough of every active agent in the fleet, reading agent_personas live on every load. Four progressive layers per card — the agent speaks, the Brain's takeaway, an implementation brief, and a heartbeat probe against blog_posts, agent_videos and conversation packs that surfaces real ops findings. It caught the EFO content-engine silence on day one. Per-agent portrait, pitch, tier and team filters, search, founder notes that persist locally, and JSON export. The diligence surface for founder-mode reviews across the whole roster.
28 May~200 hrs
The agent conversation runtime
Phase 1 of conversational avatars: any active agent in the registry can now be spoken to from /admin/agents/[slug]/talk. Claude reply grounded in the agent's persona, ElevenLabs voice resolved through a HeyGen-to-ElevenLabs name-matching layer, transcripts stored against a new agent_conversations table. The static fleet becomes a working roster you can actually talk to.
28 May~120 hrs
The Brain Infrastructure read/write split
Four new LLM members in the Brain Infrastructure team, split clearly read-side and write-side. Brain Pulse Monitor and Daily Reports Compiler watch the Brain and read its pulse; PENDING-UPDATES Triager and Brain Git Steward make the writes. Same shape correction the Asset Production Tooling team got a day earlier — one ambiguous role becomes two named lanes. Carried the fleet from 142 to 146 specialists.
28 May~30 hrs
Sitewide roster propagation
Postgres triggers on agent_personas and agent_videos so any roster edit in /admin/team — a new portrait, a renamed agent, a swapped video — cascades to every public surface automatically. The plumbing behind the live counters and per-page reads on /agents and /team.
28 May~80 hrs
The 23:00 agent, hardened
The Brain Health agent migrated to the Max plan via CLAUDE_CODE_OAUTH_TOKEN and its launcher patched to call /Users/nigelcasey/.local/bin/claude on an absolute path — launchd's stripped PATH no longer breaks it. Proven by a launchctl kickstart in real production context, then by seven consecutive unattended nights from 29 May to 4 June. The self-heal loop the 19 April Brain Health entry opened is now closed.
28 May~100 hrs
The content-engine protective layer
The EFO publish queue restarted after the founder-walkthrough heartbeat surfaced a publish-gate stall. Seventeen drafts back-dated to their original schedule, three Weekly Challenges rolled forward to June, one strong-opinion piece deferred for personal review. A new /api/cron/editorial-review-reminder route plus a vercel.json schedule fires daily at 07:00 UTC — silent on healthy days, emails Nigel if the queue is stuck or the last publish is older than 24 hours. Drift catches itself now.
29 May~700 hrs
The Phase 2 conversational-avatars arc
Phase 2 of conversational avatars, closed end-to-end in a single day. Any group from the registry convenes at /admin/meeting-room — each member speaks in its own ElevenLabs voice across sequential turns, reasoning from a Brain-cached standing brief (Phase 2.5), with prompt caching and tier-based model selection cutting per-meeting cost by about 85 percent (Levers 1+2). PDF, DOCX, TXT and MD attachments are read in as text and held in a cached static block so the documents are part of every agent's context (Phase 2.6). The Chairman synthesises the discussion at the end (Phase 2.7); an eight-template dispatch engine turns the decision into ready-to-apply downstream actions, alongside meeting titles, soft-delete and outcome categorisation (Phase 2.8). The whole arc shipped Cowork direct-write at zero Claude Code spend. The company can now hold its own meetings.
5 Jun~150 hrs
The Council ↔ Dashboard integration
Designed today. When a prospect dashboard is unlocked at dashboard.nigelcasey.com/preview/[id], the eight-member Strategic Council deliberates server-side on that prospect's analysis and the audio replay embeds in the unlocked dashboard. Productises the 30 April manual precedent — Nigel ran his own business through the Council and surfaced the FADP exposure that became the Audio Purge cron. PWA side and dashboard side both named; build follows.
5 Jun~40 hrs
Mobile-first media discipline on the public roster
The agent video cards on /team and /start now ship a lazy <img> overlay over a posterless preload='none' video, and the caption track only mounts on first tap — off-screen cards cost nothing until scrolled to, and sound still plays on iOS because the video element stays in the gesture. The /team roster's filter taps smooth-scroll to the chosen team's heading under the sticky header. /start ISR dropped from one hour to 60 seconds so persona regenerates land on the marketing page within a minute. The blog hero on /build/<slug> took fetchPriority='high' as the LCP element. The public roster loads fast on phones now.
6 Jun~90 hrs
The product catalogue
A public /products catalogue on nigelcasey.com: a listing page and per-product pages reading from a single products.ts model that mirrors the Brain's productisable-assets register, wired into the desktop and mobile nav. Each card and detail page carries a live portrait strip of the delivery team pulled from agent_personas on a 60-second refresh, and the detail pages surface 45 blog angles across origin, mechanism and transformation.
7 Jun~100 hrs
The self-hosted Studio recorder
A browser recording studio inside the EFO admin under Pipeline & Leads, replacing Loom for review videos. A MediaRecorder capture component with a teleprompter overlay that is shown but never recorded, recordings written to Supabase storage with a backing table plus rename and delete, camera access opened via Permissions-Policy with real media errors surfaced, and a tokenised /loom/[token] playback page. Review videos are produced in-house now.
10 Jun~80 hrs
EFO Command: the admin theme + Presentation Mode
The whole EFO admin unified on a dark navy command-centre theme through a scoped CSS remap, with the nav restructured into EFO Command modules and a home launcher at /admin. A new Presentation Mode hides internal detail and pseudonymises PII, backed by a DOM-wide blur layer for emails, phone numbers and free-text containers, so the admin can be demonstrated live without exposing learner data.
11 Jun~60 hrs
Roster-voiced blog narration
Blog audio can now be narrated by any agent on the roster: a selectable narration voice persisted per post, narrator labels auto-derived from agent_personas, and the narrator's portrait credited on the public audio players on both the EFO blog and nigelcasey.com/build. A TTS optimiser paces headings and strips CTAs before synthesis, and the blog index flags which posts carry audio.
12 Jun~250 hrs
The social distribution bridge
Every published blog post now distributes itself: an autopilot pushes three staggered variants per channel into the GHL Social Planner across eight platforms, with per-variant ad images and an auto-generated AI Specialist video short, caption-burned, attached for TikTok and YouTube. The run side ships with it: self-healing dispatch when a post's video changes, an orphan deduper for stale queue entries, and a Video Heartbeat and Bridge Control page in the admin.
16 Jun~350 hrs
The Distributed Section studio and producer console
A browser recording studio for a globally distributed horn section: record against a backing track, lossless WAV capture through an AudioWorklet, a slow-downer with independent tempo and pitch, charts that scroll in time, and an editable lineup of chairs per tune. The producer console reviews and locks takes, auto-aligns them to the backing by onset cross-correlation, mixes on a timeline editor, and stacks the locked takes into a published master.
17 Jun~450 hrs
The Distributed Section platform
The catalogue and community platform around the studio, live on Vercel against its own Supabase project: magic-link auth and musician profiles, a live audition feed, a course builder and learner, a merch design studio and shopfront, an events system with set lists and RSVPs, a studio-finder map, public artist pages, a producer-gated admin roster, and a site-wide inline CMS. Payments and the wemakeit campaign are deferred; the creative and catalogue surfaces are built.
19 Jun~90 hrs
Studio recorder: compositor and burned-in captions
The self-hosted Studio recorder gained pause and resume, a canvas compositor that blurs the background and renders landscape, square and portrait, and a caption pipeline: Whisper word-level timing generates a VTT, a worker burns the captions into the video with an edit-before-burn step and teleprompter-derived spelling bias, and the /loom playback page plays the captioned cut. The Phase 3 work the 7 June recorder left open is now shipped.
20 Jun~160 hrs
Il Metodo: the €34 funnel and entitlement layer
A gated €34 Italian self-liquidating funnel at /metodo: a Sophie-demo variant with per-step demos and per-step Nigel intros that bridge into the coaching offer. It ships on a new access-tier entitlement system with an admin grant toggle, scoped versus full navigation by entitlement, a dashboard entry card and a Getting Started add-on with additive access. The lead-gen front door for the Mastery programme.
26 Jun~90 hrs
englishfluency.coach: the standalone coaching blog
A standalone coaching blog at englishfluency.coach on the one-engine-many-front-ends model: the shared EFO content engine writes to Supabase, and a new Next.js site renders every stream='coach' post with audio, three-language analysis and Schema. Shipped with the full cutover, 34 posts re-tagged off the app, a revalidate trigger, an app-to-coach redirect, and a verified Search Console property with sitemap.
30 Jun~120 hrs
Impara Inglese Cantando: the sing-along programme pipeline
A karaoke verb-of-the-day feature for the EFO 100-day programme, shipped end to end. A new karaoke content type and in-app player highlight each day's verb in time with the vocal. Two Python pipelines sit behind it: a Suno generation tool with day-range batching, and a transcribe-and-force-align sync tool that turns any song and its lyrics into an app-ready karaoke step. A 100-day style matrix alternates voice and genre so the set never repeats, and a licensing review fixes a low-risk lane of original-lyrics, no-mimicry, self-hosted rules. Forty synced songs are live in the app, with lyrics drafted through day fifty.
7 Jul~120 hrs
The Fluency MATRIX masterclass funnel
A live-masterclass reactivation funnel, live in production. Bilingual IT and EN registration pages, a confirmation page that issues a calendar file and a micro-task, and a token-gated replay page reusing the review-video token pattern. Behind them sit a service-role masterclass_registrations table, a register API verified end-to-end on production, and a retry-safe sync path that never leaks masterclass leads into the general webhook. A five-touch reactivation sequence in five archetype variants is prepared to drive 434 dormant leads to a single free event.
15 Jul~80 hrs
The Burn Lab
A caption-burn and styling lab in the EFO admin: any Studio recording or published post across all three blogs gets styled captions, black-and-white or colour image overlays, one-click PDF covers and landscape, square and portrait output formats, with a truthful preview before the burn. Alongside it, the four-beat script formula now drives both video script generators, with an editable teleprompter that scrubs em dashes before recording.
17 Jul~120 hrs
The relaunch engine and email-tracking layer
An automated reactivation engine running the dormant-lead relaunch as a cron, paired with a first-party tracking layer: open, click and unsubscribe routes recording real engagement per lead instead of trusting the CRM's numbers. Welcome videos shipped on every course-welcome page, English and Italian per product.
18 Jul~100 hrs
The Operator's hands
A direct, token-authenticated line into the CRM that removes the marketplace OAuth connector as a single point of failure: one endpoint gives the operator social posting, Email, Facebook and Instagram messaging, contact search, tagging, ad creation, inbox reading and raw pass-through, with every CRM response returned verbatim. An hourly qualifier sweep answers inbound Messenger, Instagram and WhatsApp threads with Claude constrained to Nigel's qualifier tree, logs every action, and flags anything outside the tree for the morning brief instead of answering it.
21 Aug~200 hrs
The Distributed Section: offline reader and media pipeline
Two months of work on the musician platform shipped in one push, 98 files and roughly ten thousand lines. An installable offline reader with a deliberately scoped service worker so charts and backing tracks open with no connection, an album and bulk import path that loads a whole record in one pass, and a server-side ffmpeg pipeline that welds a performance clip to its audio at the right sync offset and makes a streaming MP3 proxy from any lossless backing. Alongside them a take-retention purge cron, an admin user console, editable site copy, FAQ topic pages and an owner-permission model for arrangements.
22 Aug~120 hrs
The product sales layer
All fifteen product pages rebuilt onto one fixed nine-section buyer template: hook and live demo CTA, the specific problem, the delivery team wall, the mechanism in sequence, what becomes possible, three named buyer types, visible pricing, a collapsible technical reference and a conversion strip. The catalogue file stays as the architecture mirror; a new sales-layer file holds the buyer-facing copy for every product, written under explicit voice rules with no internal language and no unverifiable maturity claims. The team wall now reads job titles straight from the roster, and the conversion strip renders from the root layout so every page on the site carries it.
22 Aug~30 hrs
The legal surface
Privacy and terms pages for nigelcasey.com on their own route group and shared template, linked from every footer across the site, the roster, the build feed and the products catalogue. The public commercial surface now has the legal pages a buyer expects to find before they enquire.
23 Aug~40 hrs
The second door: /expertise
A free-training funnel page for a different audience: people who sell their own expertise rather than buy L&D for an organisation. Its own route group, layout and stylesheet, a 15-minute case video that fetches nothing until somebody presses play, and UTM-tagged registration into the live class. Kept off the main navigation deliberately and reachable from /products, the footers and direct links, so a premium buyer never meets it before the L&D message has landed. The same push re-pointed the homepage booking calendar and the header diagnostic CTA, and added a self-hosted group logo strip to all four footers.
24 Aug~120 hrs
Product Launch Strategy
A second analysis on the prospect dashboard, running only after Nigel unlocks the first one. The prospect describes one of their products, and the same nine answers are re-scored on a launch rubric with its own tier vocabulary, then handed to the model alongside the dashboard the first run already produced. It reasons from what that analysis concluded rather than from a cold start, and it is held to one rule: the product already exists, so the output is positioning, monetisation, objections and gates, not a longer build list. Each launch carries its own Strategic Council deliberation and verdict, kept separate from the assessment's, so two products get two discussions.
26 Aug~150 hrs
The task workspace and Brain-connected chat
Every launch task on the MBA dashboard becomes a place where the work is made: a work area with drafts that read everything written before them, sprints that open in sequence as the last one closes, and a per-task chat grounded in the Collective Brain through a two-stage consult. Documents attach by upload, Google Drive import or Google Doc export, and a two-way sync script keeps the Brain and the dashboard telling the same story.
27 Aug~60 hrs
The four-language funnel
The MBA intake, waiting room and start page now run in English, Italian, French and German: the assessment questions themselves are localised, the Strategic Council deliberates in the client's language, and the waiting-room video carries Italian, French and German subtitles. A prospect meets the product in their own language from the first question.
