How Valeon Works
In November of last year I described this place as a house, and promised to keep showing you the wiring behind the walls rather than only the view from the street. That promise has held. What has changed is the house.
Almost everything load-bearing has been replaced since then. The blog was rewritten from the ground up. The audio pipeline was pulled out of a pile of build scripts and handed to a platform of its own. The author dashboard went from an idea to the centre of the whole operation. And the last of the manual steps between a finished essay and a published, narrated, syndicated post quietly disappeared.
What has not changed is the number of people living here. Valeon has published 178 posts. Every one of them was written by me.
That is the honest shape of things, and it is the reason for this post. There is no single page on this site that says what Valeon is now, how it works underneath, and what it does for the people who write on it. Scattered across eighteen months of update posts, yes. In one place, no. So this is that page — a reference for readers, for search engines, for whatever model ends up summarising this site to someone who asked about it, and for the writers I would like to hear from.
Where Valeon Stands
The archive is the least arguable evidence, so start there. 178 published posts, the first dated 6 April 2025. 69 of them this year. 173 carry a narrated audio track — 97% coverage — and the same 173 sit in the podcast feed. 1,663 minutes of reading time, a median post of 8 minutes and an average of 9.3. Eight categories, 1,787 tags, four series in use. Since the beginning, an average of 10.5 posts a month.
The platform underneath was rebuilt in May 2026: Next.js on Vercel, Convex as a reactive backend, Clerk for identity, TypeScript throughout, Tailwind and shadcn/ui over Base UI primitives for the interface. The rebuild absorbed the architectural discipline that came out of Plutarc — separation of concerns, abstraction layers that earn their keep, auditable actions, error handling that surfaces rather than swallows, and role-based access control that models real trust hierarchies.
The number I keep coming back to is the publication pipeline: roughly ten minutes before the rebuild, roughly forty seconds after it. The rebuild also moved from static rendering to SSR, which should have made things slower. It got faster anyway, because rendering, synthesis and alignment stopped waiting on each other and started running as independent jobs. The gain came from parallelism rather than from any individual step getting quicker.
If you want the current version of any of this rather than a snapshot, the stats page and the tech stack page update themselves.
The Valeon Podcast Audio Pipeline
Every essay on this site becomes an episode. The Valeon podcast is not recorded, edited or uploaded by anyone. It is a by-product of publishing, and there is exactly one human step in the entire chain: writing the prose.
Here is the whole thing, end to end.
An author hits publish. The post goes live immediately — readers can read it while the rest of the pipeline is still working, and the page upgrades itself as each job lands, with no refresh required.
The render job converts Markdown to HTML through a unified pipeline of remark and rehype plugins, handling tables, footnotes, code, heading anchors and math. Fenced code is syntax-highlighted at render time with Shiki. Equations render through MathJax as inline SVG, so nothing has to be downloaded on the client.
In parallel, the backend produces a second rendering of the same post that no reader ever sees: a speech-compatible text, built for a voice rather than an eye. This is where mathematics gets handled properly. Temml converts the TeX to MathML, and the Speech Rule Engine converts that to natural spoken English, so a listener hears "x squared plus two x" where a reader sees the equation. The two renderings stay in step with one another.
That speech text goes to VocaSync, which was originally built to be the backbone of this blog and is now a platform in its own right. VocaSync synthesises the narration and force-aligns it, returning an MP3 together with word-level timestamps keyed to both the audio and the transcript. Synthesis and alignment are chunked and processed in parallel by containerised workers with hard resource limits, retries and reactive job claims. My longest post — around 60,000 characters — used to take forty minutes to synthesise and align. It now takes about nine.
The audio and the alignment data come back to the blog. The player at the top of each post is wavesurfer.js, decoding peaks in the browser so the waveform you scrub through is the real signal rather than decoration, with hover preview and a mini player that docks to the bottom of the screen and follows you down the page. Your position is remembered between visits.
The same pipeline regenerates podcast.xml, which is picked up by YouTube, Spotify, Apple Podcasts, Amazon Music, and any third-party podcast app that speaks RSS.
Short-form clips are cut from the synthesised audio by Vizard.ai and uploaded automatically. I will be precise about the current limit of this: those clips go to my personal TikTok, and they only ever carry my own posts. Other authors' work is not clipped to my account. A dedicated Valeon channel for cross-author clips is planned, as is letting authors link their own accounts so their posts are published automatically. Both sit behind funding.
Every stage of this is observable rather than a black box. The dashboard's pipeline page shows a render card, an audio card and a podcast card per post, with stage markers for what has completed, what is running and what failed, and a retry on anything that did. A system that tells you what it is doing while it is doing it is a system you can trust and debug.
One more thing worth stating, because it explains design decisions that look excessive at current volume: the pipeline is built to handle upwards of 100 posts an hour. It currently processes the writing of one person. That gap is deliberate.
Reading and Listening at the Same Time
Word-level alignment is the reason the audio here behaves differently from a narrated article elsewhere. Since the timestamps are baked into the page as data attributes, the text and the voice know where each other are.
Highlighting lights each word as it is spoken and gently auto-scrolls to keep the current line in view. The scroll nudges rather than yanks. It is off by default, because nobody should arrive at an essay and find it turning into karaoke — you turn it on from the utilities menu in the header.
Click-to-seek is new, and it points the same alignment data in the opposite direction. With it enabled, the prose becomes a seek surface: click any word in the post and the audio jumps to the moment that word is spoken. If you drifted off two paragraphs ago, you click where you left the thread. If you are reviewing a passage and want to hear how a particular sentence lands, you click the sentence. It also lives in the utilities menu.
That menu is also where you choose your theme (dark by default, light available), your reading typeface — Inter, Lora or Merriweather — and OpenDyslexic if that serves you better. Preferences persist locally across visits.
The Author Dashboard
author.valeon.blog is a bespoke dashboard, built for this publication rather than adapted from something else. It opens on a work queue: your drafts, your submissions, anything awaiting review, and the pipelines currently running.
The editor is TipTap over ProseMirror. There is a toolbar for headings, marks, lists, links, tables and rules; a slash menu for fast inserts (/image, /file, /table, /link, /post); a bubble menu on selection; a drag handle for restructuring; and a source mode for editing the raw Markdown when that is faster. Drafts autosave as you type. An outline panel tracks your headings and lets you jump between them. Inline and block LaTeX are supported with a searchable catalogue of symbols and templates. Cmd+/ opens every shortcut and slash command in one searchable dialog.
The metadata panel holds the URL slug, cover image, excerpt, canonical URL, categories, tags, series membership and the audio toggle.
The media library stores images and documents for reuse across posts, up to 25 MB per file, in grid, list or detail views. Alt text does double duty as the visible caption and the screen-reader description. Every file serves through a stable /m/{id} proxy, so an embedded URL keeps working forever regardless of what moves underneath it.
AI assistance is optional and always a starting point rather than an output. It will generate a cover image from a subject and one of eight editorial styles, draft a one-sentence excerpt from your body text, and describe your cover image for alt text. You review and edit everything before it saves.
Cross-post references link to the post rather than to its URL, stored as a valeon:post:{id} reference. Rename a post, move it, change its slug — the link survives.
Every author gets an @valeon.blog mailbox with a custom inbox in the dashboard: read, reply, archive, mark unread, and optionally forward to your personal email. Readers reach you without you publishing a personal address.
Your author profile carries an avatar, bio, links for website, X, GitHub, LinkedIn, Bluesky and Instagram, and up to 25 external publications — books, papers, talks, work published elsewhere. It renders as a public page at /authors/{your-slug} with a clean index of your essays and its own RSS feed, designed to be portable enough to link from a portfolio.
Account security covers passkeys for passwordless sign-in, password management, and a list of active sessions you can revoke device by device. API keys are created and revoked from the same place, which is what the Obsidian plugin authenticates with.
Data export produces a complete archive of your posts, drafts and media — Markdown files with YAML frontmatter, media bundled alongside. This is a matter of principle rather than a feature. Your writing arrives in a portable format and it leaves in one.
Behind all of that sits a role model — author, editor, admin — with a direct-publish permission that lets trusted authors skip review entirely, and an append-only audit log of consequential actions.
Publishing from Obsidian, Without Ever Opening a Browser
If you already write in Obsidian, the Valeon plugin is in the community catalogue, and it is the detail I am most pleased with in the whole platform.
Install it, create an API key in the dashboard, paste it in, test the connection. Sync the template first, which pulls down the schema and taxonomy so your vault knows the real categories and series. Your vault layout is a folder per post — YYYY-MM-DD-slug/post.md with an assets folder alongside it.
After that the daily flow is three commands: new post, lint, publish. Frontmatter, body and embedded images all sync in one step. The AI helpers are available from the command palette, so covers, excerpts and alt text never require the dashboard either.
Sync runs both directions. You can pull a published post back down, pull a draft buffer, or pull metadata alone. A fresh vault on a new machine can restore your entire archive. There are bulk and recovery commands for syncing, restoring and reconciling the vault, and conflict handling with three explicit options — back up, force, or cancel — when local and remote have both changed.
It works on iOS and Android with every command available, and pairs with the community Linter plugin for frontmatter hygiene.
No git. No terminal. No frontmatter written by hand, no asset folders to manage, no browser between the thought and the page.
What Funds This, and What That Costs
Valeon is a non-profit project. It generates no revenue, and that is a permanent decision rather than a stage it is passing through. It will never carry advertising. It is open to sponsorship. It is funded entirely by my personal income and by revenue from the other products under the Valeon umbrella — Plutarc, VocaSync, gramatic, and the rest.
That economics explains the automation honestly. Automation is the cheap path, and the cheap path is the one available to me. Human editorial review — of the clips in particular, where an algorithm is cutting without judging — is on the roadmap and gated behind having the funds to bring editors on for that step. I would rather say that plainly than dress a constraint up as a philosophy.
The terms for writers follow from the same posture. You retain copyright at all times. Accepted work is published under a 30-day exclusive digital licence followed by a perpetual non-exclusive one, so after the exclusive window you are free to syndicate anywhere with a canonical link back. Claims get sourced and linked inline. AI assistance, sponsorships and financial ties get disclosed. No unsourced medical, financial or legal advice.
Write With Us
Valeon publishes long-form essays by invited authors, across philosophy, markets, technology, science, psychology, culture and the craft of making things. Essays run 1,200 to 3,000 words, features 3,000 to 6,000. Poetry is exempt from both.
You apply with a sample of your work through the How to Contribute page. Applications land in an editor's queue and get a response within seven to ten days. If it is a fit, an emailed invitation sets up your authoring account — no sign-up form, no password created before we have actually spoken. From there you write in the browser or in your vault, submit for review, address inline comments, and publish.
On the day your account is active you get all of it: the dashboard, the editor, the media library, the mailbox, a public author page with its own feed, a narrated audio track for every essay, distribution to four podcast platforms, and an export button that hands your archive back whenever you want it. Nothing is held back for a later tier, because there are no tiers.
The author documentation is public, so you can read exactly what publishing here involves before deciding whether to apply.
The house is built. The wiring is in the walls and the pipeline runs itself. There are simply more rooms in it than there are people, and that is the only part of this I cannot solve by writing more code.