Sunday Latte: Building Andy’s Café

A café already open

This Sunday Latte is a little different. Instead of explaining one piece of artificial intelligence, it looks at the place serving the explanation: Andy's Café itself. The project began with a simple personal need. A difficult subject was easier to understand while walking, doing chores, or repairing something than while staring at another screen. Turning a careful explanation into a podcast made the knowledge portable.

That private tool has become a public, multilingual café. As this episode is published, the core shelf contains eleven shared editions in five languages: English, German, Italian, Spanish, and France French. They cover basic concepts, relationships among concepts, larger questions, and two weeks of AI news. This project update is an extra English edition, not the announcement of a rigid new format.

The surprising part is not that synthetic speech can read a document. The interesting part is the complete path: choosing what deserves attention, establishing the facts, writing for the ear in several languages, finding a voice that actually belongs in each language, making the result easy to discover, and operating the whole thing without turning it into another machine for tracking or manipulating people.

Four cups, four jobs

The menu has four recurring shapes. A Double Espresso explains one term with high density. Coffee and Cake connects a few related ideas. A Sunday Latte takes time with a larger question. The Morning Paper selects and explains current developments. Those names are playful, but the distinctions matter: a short definition, a comparison, a patient exploration, and a news edition ask different things of both writer and listener.

The early catalogue was organized as a small matrix so every format could be exercised in four languages. That matrix did its job, and then stopped being a target. It was never a constitution. A useful podcast does not become better because a spreadsheet contains the right number of green cells.

The next editorial direction is therefore flexible. More Espressos make sense because they are short, memorable, and form a useful vocabulary shelf. Morning Papers matter because the field changes quickly. Sunday Lattes and Coffee and Cakes earn their place when a subject genuinely needs their shape. Cadence should follow what is useful and sustainable, not a promise made before anyone has listened.

One factual core, several languages

A multilingual edition does not begin by writing English prose and replacing each sentence with another language. It begins with a shared factual core: sources, numbers, dates, uncertainty, and the distinctions that must survive. Each language edition is then written for its own grammar, rhythm, terminology, and listener.

That difference becomes obvious in speech. A phrase that looks tolerable on a page can sound absurd when spoken. A technical term may be familiar in English but need a short explanation in German, Italian, Spanish, or French. Names and abbreviations can survive transcription while still sounding wrong to a native listener. Editorial review and listening therefore remain separate from mechanical checks.

French illustrates the approach. It did not wait for the entire historical catalogue to be recreated before the first useful episode went live. A France-French voice and one Espresso established a real public shelf. The remaining editions could then follow through the same factual and production path. The choice is still provisional: competent listeners can change it. A language launch is a beginning, not a declaration that pronunciation research has ended.

Voices are language-specific

There is no single best text-to-speech system in the abstract. A model can sound excellent in German and Italian, acceptable in English, and completely wrong in Spanish. That happened here: an early Spanish voice spoke intelligible Spanish with the unmistakable cadence of an English native speaker. The problem was not rescued by admiring a benchmark. The voice was rejected, a different engine took over Spanish, and production continued.

English, German, Italian, and France French currently use language-matched reference voices with a larger cloning model. Spanish uses a different, smaller system whose selected voice sounds much more naturally Spanish. The short café host is another voice again. Convenience would favour one engine for everything, but listeners do not owe a production pipeline that convenience.

The mechanical side is intentionally resumable. Editorial paragraphs are separate speech requests. If one sentence stutters, drops a date, or mangles a name, that paragraph can be replaced without recreating twenty minutes of clean audio. Independent editions can render in parallel on temporary graphics processors, while packaging and feed publication remain one-at-a-time operations. That gives the project speed without allowing two workers to overwrite the same public catalogue.

From script to phone

Once a script is final, the rest of the path is deliberately plain. Numbered chapter files become matching audio chapters. A short piano signature, a quiet transition, and the café host are composed around the narration. The packager creates one chaptered audio file. A separate release step gives it a stable identity, copies it to the public shelf, and updates the language feed.

That separation matters. Rendering is expensive but reversible. Packaging creates an inert draft. Publication is the moment a listener can receive a change, so it remains explicit and serialized. A failed experiment cannot silently replace an episode that already works.

The delivery technology is RSS, a wonderfully unglamorous piece of the open web. Each language has its own feed and can be added to an ordinary podcast app. The finished audio and feed live as static files. There is no account to open, no proprietary player to install, and no database required to serve an episode. If a directory disappears, the feed still exists.

A website for humans and agents

The public website has two jobs that pull in opposite directions. A busy human should understand the offer, choose a language, play or subscribe, and leave within seconds. A curious visitor should also be able to browse the catalogue, read transcripts, compare language editions, and understand the editorial idea.

Agents are visitors too. Welcoming them does not require a theatrical agent portal or a new application programming interface. It requires the ordinary web to be unusually legible: stable links, semantic HTML, RSS discovery, a sitemap, plain-text transcripts, and a small machine-readable catalogue. The same public corpus can serve a screen reader, a search engine, a research tool, and a person on a phone.

Transcripts are now a deliberate part of the product. They make an audio explanation searchable, quotable, accessible, and reusable. They also make source quality more important, because machine visitors can copy errors as efficiently as truths. The aim is to be easy to use, not easy to exploit: sensible caching and rate limits may protect the service, while the content itself remains openly reachable.

Growth without extraction

Distribution is now one of the hard problems. A technically perfect feed that nobody finds teaches nobody. The ambition is therefore large: descriptive episode titles, podcast-directory listings, language-specific discovery, searchable transcripts, useful links from relevant communities, and a catalogue worth returning to. French is not only another audio track; it is another audience, with its own vocabulary and places of discovery.

But growth is not permission to recreate the worst parts of the advertising internet. Andy's Café should not trick people into clicking, build behavioural profiles, rent attention, or send autonomous spam. It can be ambitious without being manipulative. The durable growth loop is simpler: publish something genuinely useful, make its subject obvious, let people and agents link to it, and make subscribing painless.

Measurement still helps. The useful questions are modest: which language and episode was requested, roughly where the request's network exit was located, how much was downloaded, and what the production effort was. Those facts can be aggregated without retaining an address, a complete browser signature, a cookie, or a minute-by-minute journey. Counts should guide decisions, not become surveillance dressed as product insight.

The content problem

Distribution is difficult, but choosing what deserves distribution may be harder. The public internet increasingly repeats itself. A press release becomes a news story, the story becomes a summary, the summary becomes a model answer, and the answer becomes another article. Counting mentions would reward the loop rather than reveal importance.

The Café needs a different discipline. Original papers, official documents, first-hand statements, and careful reporting should anchor the factual ledger. Collection can be automated; judgment cannot be reduced to popularity. A source collector is an editorial inbox, not an oracle. It should help find material while preserving where a claim came from and how uncertain it is.

This connects editorial judgment with the degradation of public information. The answer is not nostalgia for a purely human web, nor blind faith in a purely automated one. It is traceability, comparison, proportion, and the willingness to leave a fashionable story out. The project exists to help listeners understand, not to fill airtime.

What comes next

Several tracks can now move at once. The existing catalogue can keep expanding in French. More Double Espressos can build a useful vocabulary. The next Morning Paper can collect sources while the news week unfolds. Podcast directories can improve discovery without stopping language work. The website can become a faster catalogue, and the transcript corpus can become easier for both humans and agents to navigate.

There are open questions rather than gates. Which language offers the next combination of reach, editorial feasibility, and convincing speech? Does the French narrator remain comfortable over many episodes? How should the four formats share a weekly rhythm? Which titles actually help people discover an explanation? Can privacy-preserving aggregate measurements reveal enough without inviting extraction? Would a phone application add independence, or merely duplicate what RSS already does well?

Some ideas are deliberately waiting. Personal podcasts generated on demand could be valuable, but introduce product, safety, and prompt-injection problems. An autonomous maintenance agent sounds attractive, but has not shown that it would do more than focused tests, live probes, and human escalation. Hardware may move home one day, but temporary capacity works today.

The governing phrase is simple: nothing is carved in stone. That does not mean drifting without a plan. It means treating every voice, format, tool, schedule, and distribution tactic as a choice that must keep earning its place. Andy's Café is already open. The next task is to serve more good cups, help more people find them, and keep tasting what the machines brew.