Lloyd D’Souza

The Dossier

Channel Intelligence is an editorial video-intelligence dossier built for Condé Nast editorial and audience leads. Pick a title — GQ, Vogue, Bon Appétit, WIRED, The New Yorker and the rest of the U.S. portfolio — choose a channel or the title’s whole U.S. network, set a window, and it compiles a magazine-style briefing on how that channel is actually performing: the topline, the videos that carried the window, who watched and where they came from, what the Shorts business is doing, and how the owned TikTok and Instagram handles are moving alongside. It is designed to read like a briefing rather than a BI tool, and every number in it is one the leads can take into a programming or budget decision without caveats, because each one reconciles to the source it came from.

A Dossier

GQ · YouTubeScreen

Configuration & subject channel

Screen

B Topline

GQ · 90 daysScreen

Topline, composition & lifecycle

Screen

The Plug-ins

The dossier is modular. The YouTube layer carries the core reads — Performance, Shorts, Audience, Competitors, Sponsorships — and within each, a set of plug-in analyses expands on demand: Top Assets, Topic Terrain, Title Grammar, Hook Study, Length Signal, Cadence and Engagement Mapping, each an AI-drafted editorial read over the window’s uploads, reviewed by an editor. Beyond YouTube sit the owned-social layer (TikTok and Instagram from the company’s Dash Hudson data), Open Web listening for brand mentions, and a multi-dossier desk so several channels or windows can be open side by side and compared. Editorial context lives as data, not caption copy: a registry encodes channel lineage (a migrated channel’s history rides inside its successor), which channels a title offers, and which handles count — so the numbers change, rather than being explained.

C Plug-in

Top AssetsScreen

Top Assets — the head and the tail

Screen

D Read

Audience · VisualScreen

Audience — geography & cohorts

Screen

E Owned social

@gq · TikTokScreen

TikTok — profile & trajectory

Screen

F Owned social

Dash HudsonScreen

TikTok — what travels

Screen

Under the Hood

Underneath, a curated data layer does the work. A nightly job rebuilds two dozen tables in the company’s Databricks lakehouse from the raw YouTube Analytics and social extracts — channel-by-day, video-by-day, geography, demographics, traffic, devices, lifecycle, monetized ad impressions, owned-social posts — on a perimeter that matches what the leads select, with every table validated against YouTube Studio or Dash Hudson before it ships. Two decisions set its character. Engaged views, not the platform’s redefined “views”, are the house metric, with a toggle to see plays as reach; and year-over-year comparisons are withheld rather than shown when the two windows sit on different definitions. The app itself is a FastAPI service and a Next.js front end deployed as a single Databricks App, so the people reading it sign in with their company identity and the data never leaves the lakehouse.

G Pipeline

Databricks · nightly refreshScreen

Databricks job — the curated refresh, table by table

Screen

The Analyst

The Analyst is the dossier’s resident analyst — a side drawer you can talk to. It reads the dossiers open on your desk and, when a question reaches past them, pulls exactly what it needs through the same vetted queries the dashboard runs: any window, any channel, year-over-year, a specific video found by title and read against its peers, a ranked leaderboard for a custom period, monetized impressions, owned-social activity. It can also look outward — exact public counts for any channel on YouTube, and web search with sources for what is happening in the landscape — under a strict rule that curated, public and web figures are never blended and every figure is labelled with where it came from. It will not estimate revenue, because the booked number is not in its reach and an estimate would not be exact. Its value is less the answers than the time they take: a question that used to mean an export, a pivot and an afternoon is now a sentence, with the work shown, the numbers pulled live and the conversation exportable as a record.

H The Analyst

Ask the dataScreen

The Analyst — a question, answered with the work shown

Screen

The Edition

Folio is a cross-brand editorial discovery app for Condé Nast, built as a working prototype and running at condenastfolio.com. It gathers what fourteen mastheads publish — Vogue, The New Yorker, Vanity Fair, WIRED, GQ, Architectural Digest, Condé Nast Traveler, Bon Appétit, Ars Technica, Teen Vogue, Pitchfork, Glamour, Allure and Epicurious — and presents it the way a magazine would rather than the way a platform does: as a numbered, dated edition of twenty-four pieces that ends. There is no infinite scroll, no engagement loop and no algorithmic mystery. Every headline is set in its own masthead’s face, every piece opens on the publisher’s own site so the page view and the subscription stay with the brand, and the whole thing runs on a warm near-black ground that reads as a lookbook rather than a website. It installs to a phone’s home screen and, above 1024 pixels, lays the same screens out for a desk.

A Edition

Edition No. 271Desk

The desk — the lead row, the pull-quote card, the grid beneath

Screen

The Plug-ins

Several parts plug into that edition. An ingest job reads each masthead’s feeds twice a day and files roughly 21,000 articles, galleries and recipes, 38,000 videos and 3,000 podcast episodes into one catalogue, refusing coupon pages at the door. An enrichment job inside Condé Nast’s Databricks lakehouse reads each article’s body where it already lives, and writes back only derivatives — a word count, a pull quote in the writer’s own voice, three editorial labels and an embedding — so no article text ever leaves the workspace. Search runs a hybrid of keyword and vector retrieval with concept expansion, so “quiet luxury” finds Phoebe Philo and stealth wealth across six titles, most of them containing neither word. Studio holds the video and audio: YouTube clips behind the app’s own transport, podcasts playing in-app at the publisher’s own measured URL so every listen is still counted. Readers sign in with an email and a password, keep tearsheets of saved pieces and playlists of things to watch and listen to, and can opt into the Morning and Evening Editions — two emails a day cut from the edition by rules an editor would recognise, plus a Sunday Edition that looks back at the week.

B Search

“Quiet luxury”Desk

Search — forty pieces from eight titles, the concepts read as a sentence

Screen

C

The Edition

Phone

D

Studio — watch

Phone

E

Search as a concept

Phone

F

Studio — listen

Phone

How It Works

How it works, mechanically, is a chain that runs on time. A small Cloudflare Worker fires three scheduled runs a day: the morning and evening ingests at 8am and 6pm New York, and a media run at 4pm. Each run writes to a Postgres database with vector search, embeds what is new, rebuilds the site so the front page is current to the minute, and, when the switch is on, sends the newsletter to everyone who opted in. The app is an Expo project deployed on Cloudflare Pages; the catalogue is baked into the bundle so nothing ships a database credential to a reader, and the few things that must be live — search, a title’s deeper archive, the curator’s conversation — run as serverless functions. Everything is covered by 500 offline tests that run without keys or a network, which means the whole system can be checked in a meeting.

The Curator

The Curator is the part no one else can build, and it is where the value sits. A reader trains one: they brief it with a subject, or seed it from a tearsheet they have already assembled, and then refine it a round at a time against six pull quotes drawn from the archive around its moving centre of taste — picking the ones that are right, passing on the ones that aren’t. From four picks it has a readable profile on five editorial dials: treatment, depth, stance, vintage and register. A finished curator offers three things: Read, the pieces it holds; Refine, more rounds; and Chat, an agent built from its taste that answers questions by searching the archive and ranking what it finds by relevance first and the curator’s dials second, surfacing real pieces the reader can open rather than describing them. The point is that taste crosses mastheads where brands cannot. A curator trained on one person’s fashion tearsheet will find the WIRED piece and the New Yorker essay that belong beside it, and that is a reading of fourteen titles that no single brand’s homepage, newsletter or recommendation engine could produce. For a company whose titles compete for the same attention, a curator is the first thing that makes them read as one.

G Refine

Round 7 · SS27Desk

Refine — a round of six, the dials taking shape

Screen

H Chat

Ask the curatorDesk

Chat — answers drawn from real pieces in the archive

Screen

The Hall

Peripatos Hall is a private studio I built for writing and thinking alongside AI. It takes its name from the peripatos, the covered walkway at Aristotle’s Lyceum where teacher and students worked through ideas on foot, in conversation. The Hall is a single web app drawn as an architect’s sketch of a small campus: a central spire, two scholars’ towers, a domed creative hub and a library. Each building is a room, and behind every door is a collaborator modelled on a historical figure, with a defined role, a distinct voice and access to a shared, growing library. It’s where I develop my screenplays and keep the research that feeds them.

A The Hall

Peripatos HallScreen

The Hall — each building is a room

Screen

The Collaborators

The Hall is made of a set of agents and tools, each with one job. Aurelius, in the spire, is the Hall’s orchestrator: ask him what the Hall knows about something and he searches the library and answers with a synthesis. The two towers hold the scholarship written by the Hall’s resident thinkers: Seneca (philosophy), Augustine (theology), Gibbon (history), Oppenheimer (physics) and Darwin (the sciences) in the Intellectual Pod, and Hamilton, Kennan and Sherman (politics, foreign policy and military history) in the Geopolitics Pod. The Library is run by Eratosthenes, who answers questions by connecting documents across fields, and Dewey, who maps each collection into a domain summary. The Creative Hub is a writers’ room. Scheherazade conducts it, bringing in Homer for myth and archetype, Asimov for technology and speculation, Tolkien for world-building and Stoker for dread. Salinger writes every page, and Perkins edits. Around the room sit the production tools: Panel Review, which produces industry-style script coverage; Compile, which stitches sequences into a full draft; a Perkins full pass for whole-script edits; and Writer Profiles, which learn a writer’s voice from their own samples so the drafts sound like them.

B Aurelius

The spireScreen

Aurelius — synthesising what the Hall knows

Screen

C Intellectual Pod

Gibbon’s shelfScreen

The history tower — Gibbon’s shelf

Screen

D The Library

EratosthenesScreen

The Library — Eratosthenes connecting theology and history

Screen

E Creative Hub

The domeScreen

The Creative Hub — projects and production tools

Screen

Under the Drawing

Under the drawing, the Hall is a React front end and a Python (FastAPI) back end running on a Mac mini at home, reachable from anywhere through a private sign-in. Every collaborator runs on Anthropic’s Claude, and the model fits the job. Claude Opus 4.6 does the craft work (the writers’ room, Salinger’s drafting, Perkins’ edits). Claude Opus 5.5 is the Panel’s deliberately different “outside reader”. Sonnet handles Aurelius’s and Eratosthenes’ judgment calls, and Haiku does high-volume work such as Dewey’s summaries. The library is plain Markdown that’s indexed automatically as documents arrive, so every agent can search it, cite it and build on it. In the Workshop I load a project, talk through a scene, and Scheherazade gathers each specialist’s view. Nothing gets drafted until I say so. Then Salinger writes into a live preview in Fountain screenplay format, with page counts calibrated to industry pagination. The rule underneath it all is that the voice on the page should enrich a scene that’s really happening on screen, never replace it.

F Architecture

Orchestration flowDiagram

Architecture — the orchestration flow

Diagram

G The Workshop

Midlife VampireScreen

The writers’ room at work — the conversation on the right, Salinger’s draft live on the left

Screen

The Walking School

The long-term goal is for the Hall to become a true walking school: a place where scholarship and story feed each other. I want the scholars back at work as autonomous researchers, so that the history behind a vampire’s legend, or the politics of an imagined empire, becomes source material the writers’ room can draw on directly, and the questions a story raises send the scholars off to research again. Over time the library becomes the Hall’s long memory. Writer Profiles let every collaborator write in my voice rather than a generic one. A project can travel under one roof from a first idea to a researched world, a drafted screenplay, professional coverage and a sharper rewrite. Through all of it, the author stays the one walking the path and the agents walk beside.

H Panel Review

Coverage · Opus 5.5Screen

Panel Review — industry-style coverage

Screen

I Writer Profile

Personal voiceScreen

Writer Profiles — the Hall learns your voice

Screen

In Action

▶ Walkthrough

Peripatos Hall · alphaVimeo

Peripatos Hall — agents collaborating on a scene

Video

The Desk

Aither Horizon is a private intelligence desk for world politics, built for one reader. It puts the whole world on a single living map and asks one question of everything it shows: is this a signal, or is it noise? Every two hours it reads 75 trusted news outlets and the GDELT event stream, which machine-codes news reports from around the world. It sorts what it finds into regions (the Americas, Asia, the UK and Europe, the Middle East, Africa) and into the theaters that matter most right now, among them the Iran war, Russia–Ukraine, China’s sphere of influence and US politics. Reported events rise from the map as spikes and reporting volume shapes the terrain, so the places that matter stand out at a glance. You can move from the whole globe down to a single city and every panel follows. Each item keeps its provenance: who reported it and when, when it happened, and when the desk saw it. State-owned and state-funded outlets are labelled as such.

A Watch

Globe · last 24 hScreen

Watch — spikes are reported events, the terrain is reporting volume

Screen

B Region

United StatesScreen

Region view — the wire and hotspots narrowed to the US

Screen

C City

LondonScreen

City view — London, and an AI reading of what actually happened

Screen

The Modules

The desk is built as a set of modules, each a different lens on the same map. Watch is the live view: a ranked “what matters now” list, a wire of stories grouped across outlets, hotspots ranked by how much action is reported there, and live news channels from broadcasters around the world, playing over the map. Power shows who holds power in each of the world’s 196 countries: the head of state and the head of government, who controls the legislature seat by seat, which institutions can check them, and a short reading of where power really sits. Every profile has been checked against primary sources. Its Election Watch follows upcoming national elections with key dates, polls and what is at stake, and its Military view maps defence spending, forces, nuclear arsenals, alliances, bases and troops deployed abroad. Economy puts each country’s growth, inflation, debt, currency, interest rates and trade partners on the same globe, drawing on the IMF, the World Bank, central banks and UN trade data. Two modules are still to come: Library, a store for articles, documents and notes, and Theory.

D Video

Live channelsScreen

Video — live channels over the map, state-funded ones labelled

Screen

E Military

Power · United StatesScreen

Power · Military — US forces in 51 countries

Screen

F Election Watch

Brazil · runoffScreen

Election Watch — Brazil’s runoff, what is at stake, fact-checked

Screen

G Economy

Germany · IMFScreen

Economy — Germany’s figures, markets and trade partners

Screen

H Power

United KingdomScreen

Power — seats shaded by role, the 326 needed to govern, and the institutions that can check the government

Screen

Under the Hood

Under the hood, the desk keeps three things apart: the evidence, the machine’s reading of it, and what the reader decides. GDELT is treated only as a measure of attention, showing where the world is talking, and never as proof that something happened. A place becomes an event on the map only when an AI reading of its articles confirms something important happened there, or when a trusted outlet reports from nearby. AI models tag each article with its theater, the countries it concerns, the type of event and an importance score from 1 to 5, and they group coverage of the same story across outlets. On request, Claude writes a cited brief for any region, theater or place, with every claim linked to the article it came from. Nothing is overwritten: each profile, election record and reading is saved as a new version, with its sources and the reason for the change. Anything not yet checked against a primary source is marked “To verify”. The fact-check of every Power reading went through 1,390 individual claims.

Theory

The long-term goal is Theory: to use all of this to form, test and revise a political theory in the open. Positions will be written down as explicit claims, each tied to the evidence that supports it, the objections against it and every revision it has gone through, so the thinking itself becomes an inspectable record rather than a set of impressions. A “think tank” of AI roles (a historian, an empirical analyst, a red team and a synthesist) will argue for and against a chosen claim, citing the Library and the live record. Claims that make predictions will be scored against what actually happens on the map. The machine can propose a change of mind, but only the reader can adopt one. Watch, Power and Economy are the instruments; Theory is what they are for.

The Game

Lone Wolf is a full digital edition of Flight from the Dark (1984), the first of Joe Dever and Gary Chalk’s legendary choose-your-path gamebooks — rebuilt as a premium, mobile-first web game. All 350 sections of the original adventure are playable: you are Silent Wolf, last of the Kai, racing across Sommerlund to warn the King while the Darklords’ armies burn everything behind you. The adaptation is faithful to the book — every choice, fight, death, and secret is where Dever put it — but it plays like a modern game: a cinematic “story so far” intro, scene-setting art as you cross into each of the sixteen regions, an ambient score that shifts with the mood of the land, and a two-tap die for the blind “Roll for Fate” moments that give the series its tension. It runs in any browser, fills a phone screen edge to edge, and works offline. (Play it here.)

A

Title screen

Phone

B

The story so far

Phone

C

Entering a new region

Phone

The Build

The build is deliberately tiny. The entire engine is one hand-rolled HTML file — roughly 2,400 lines of vanilla JavaScript and CSS, zero frameworks, zero dependencies — and each book compiles into a single self-contained game file. A Python pipeline does the heavy lifting: a parser reads the canonical Project Aon text and converts it into structured data (sections, choices, combats, items, conditions), a curation layer layers on hand-made fixes and story-state flags, and an audit tool enforces an integrity contract — all 350 sections reachable, all 555 choice-edges verified against the book’s flow graph, every random outcome routable — and refuses to ship anything less than a clean pass. Art and music flow through the same philosophy: drop-folder tools downscale and compress each asset (watercolor illustrations, enemy plates, discipline cards, zone backdrops, eight ambient tracks) and wire them to the game by filename convention. The whole project was built in an AI-collaborative workflow — design, code, art direction, and playtesting iterated daily with Claude — with every change audited, verified in a live preview, and deployed to the edge on Cloudflare Workers in minutes.

In Play

In play, the game automates the bookkeeping and preserves the choices. The original books made you track everything on a paper Action Chart; here, a live ledger announces every wound, meal, theft, and treasure with visible arithmetic, while the genuine Combat Results Table runs the fights — enemy art plates, condition chips that explain every modifier (Mindblast +2, fighting blind in the dark, a Vordak’s Mindforce sapping your skill from round two), damage numbers flying off the endurance boxes, and a SLAIN banner when a foe falls. What the game never automates is you: which weapon you ready (your Weaponskill bonus only counts if you’re actually wielding the blade you mastered), whether to risk the Graveyard of the Ancients, whether to trust the hooded ferryman. Kai Disciplines gate paths exactly as the book intends — and when you lack one, the game shows you what you’re missing, planting the seed for the next run.

D

Choosing Kai Disciplines

Phone

E

A section — ledger, illustration, choices

Phone

F

Combat — enemy plate, condition chips, the die

Phone

Built to Replay

It’s built to be replayed — and to grow. Progress autosaves continuously, with named save files alongside. Finishing the book earns an illuminated epilogue: enemies felled, most-used weapon and discipline, and a collection report that shows the key items you recovered in green and the ones you walked past in red, with the same treatment for zones visited — a quiet dare to go back in. A rogue-like replay mode lets you run the book again with your veteran character (no double-dipping on unique treasures), and a standalone Arena mode distills the combat engine into a pure gauntlet: twenty escalating foes, a new Kai skill between waves, and a champion’s roster if you survive. The engine was designed from day one as a saga shell — Books 2 through 5 of the Kai series slot into the same pipeline, same engine, same save-carry-over — so Fire on the Water is a data import away, not a rewrite.

G

Book complete — the collection report

Phone

H

Arena — Champion of the Arena

Phone

A free, non-commercial fan work. Text © Joe Dever; original illustrations © Gary Chalk; the book text is distributed for non-commercial use by the volunteers of Project Aon. This edition’s interface, art, and music are original to the project.