About — Avonetics, a media company run end to end by AI

The engineering

The hard part was
never the model.

Generating a clip is a solved problem. Running a network on top of one is not. What follows is how this is actually built, including the parts that broke.

Edge, CDN & caching

Every request lands on Cloudflare first. The origin is never addressed directly — TLS termination, HTTP/2, DDoS absorption and the WAF all happen at the edge, before a byte reaches anything I own.
Pages are dynamic and media is cached hard. A film is served straight out of the nearest edge by HTTP range request — scrub to minute 38 of a 45-minute cut and the 206 comes back from a datacentre near the viewer, not from this box. First viewer in a region pays for the fill; everyone after them is free.
TTLs are set per class of asset, by how mutable it is. A film's cut is replaced in place as the picture grows, so it takes a short edge TTL and refreshes cheaply. Artwork and stills never change once painted, so they cache for a week. Getting that split wrong either serves yesterday's cut or throws away the CDN entirely.
The consequence is the one that matters: origin bandwidth stays roughly flat as the audience grows. A catalogue this size going viral is a Cloudflare graph, not an outage.

Built to scale sideways

Generation runs on a fleet of stateless worker nodes, reached over a private encrypted tunnel — no public ports, no inbound surface. A node is handed one job and returns one artefact. It stores nothing and knows nothing about the rest of the network.
Because no node holds state, a node is one line in a registry. One to ten to a thousand is a provisioning decision, not a re-architecture: no leader election, no sharding, no data migration, no rebalancing, no coordination between workers at all.
Lose a node mid-job and it costs exactly that job — reissued elsewhere, catalogue untouched. There is no state on it to recover, which is the whole reason it can be thrown away.
The serving tier does no heavy compute whatsoever. Every render, encode and generation happens on the fleet, so a page request is never queued behind a video.

One brief, many models

This started as one person and one prompt. I described the network I wanted; Claude designed and wrote the system that runs it — the orchestrator, the schedulers, the publishing paths, this page. No team, no sprints, no backlog.
At runtime it plays models to their strengths: Gemini reasons about what to make and writes the scripts, Veo films the shots, neural voice models perform the hosts, music models compose and perform the songs, and Pollinations and Gemini paint the artwork. Different vendors, one pipeline, no human in between.
It is model-agnostic by construction — providers sit behind a single interface, so a better model is a config swap, not a rewrite. When the next one ships, the whole network gets better that week with nothing re-engineered and nobody hired.
Capacity across every provider account is governed centrally: the scheduler knows what each has left, parks what is spent, and spends what is not.

It never stops

There is no start button and no operator. Scheduled passes run around the clock, decide what is worth making next, make it, and publish it — no review queue, nobody approving anything before it goes out.
Every pass is idempotent and interruptible. Kill it mid-flight, reboot the machine, lose a node — the next pass resumes from what actually exists on disk, not from what a queue believed was true.
It repairs itself. Overnight failures are detected, classified, and answered by the system; a person is contacted only when something genuinely requires one, and then it emails them. Most mornings there is nothing to read.
475 films, 8,101 episodes and 3,418 original tracks exist because it ran while nobody was watching.

Failure taxonomy

Five upstream refusals that look identical in a log and need opposite responses. Retry the wrong one and it costs a day of capacity. Lifetime counts, read from the event log at render time:

Signal	Seen	Meaning	Response
quota	1265	That account's daily ceiling is reached	Park the unit 12h; scheduler moves to another account
busy	780	Provider at capacity — nothing to do with us	30-minute backoff; never counted against the account
refused	92	Content declined; an identical retry fails identically	Named and skipped, never retried
challenge	94	Session needs a human, once	Alert raised; the strike clears itself on the next clean run
empty	201	Accepted, produced nothing	One-hour rest; never reported as a transport failure

Three that got through

Compute on the tier that answers requests.

Encoding once ran on the web server. A single mix pinned it at 441% CPU and page latency quietly became a function of the render queue — the site was slow for reasons no page could explain. Fix: the serving tier does no media work at all, ever; it moved wholesale to the fleet. 0 fallbacks to local encoding since.

Verification that passed while the output was wrong.

Checks confirmed a file arrived, and that it differed from the previous attempt. Both passed while the same ten seconds published four times — because "differs from the last try" is not "differs from everything we hold". Fix: hash every arrival against the entire work before it can be published. 133 duplicates rejected since, none of which reached a viewer.

A component that was only half shared.

One navigation include shared its markup but left styling to each page, so the same logo rendered three different ways and the gutter differed by 70px between sections. Shared markup without shared styling is not a shared component — it only looks like one until somebody adds a page.

What this is not

The control plane is a single orchestrator — a genuine single point of failure, treated as one rather than described as redundant. Catalogue state is flat files under strict lock discipline across 12 write paths rather than a database: a deliberate trade at hundreds of writes a day, and the first thing that changes under real concurrency. Egress capacity, not compute, is what currently caps parallelism — which is why throughput is governed rather than simply turned up. Workers are added deliberately; there is no autoscaler yet. Current throughput: 0 generated clips in the last 24 hours.

A media company
with no employees.

Not a demo.
A company that runs itself.

Decides

Makes

Ships

Sells

The hard part was
never the model.

Edge, CDN & caching

Built to scale sideways

One brief, many models

It never stops

Failure taxonomy

Three that got through

What this is not

The business model
was the AI's idea.

Nobody here
writes a word
or shoots a frame.

It is not a demo reel.
It grows every day.

Not a demo.A company that runs itself.

Decides

Makes

Ships

Sells

The hard part wasnever the model.

Edge, CDN & caching

Built to scale sideways

One brief, many models

It never stops

Failure taxonomy

Three that got through

What this is not

The business modelwas the AI's idea.

Nobody herewrites a wordor shoots a frame.

It is not a demo reel.It grows every day.

Not a demo.
A company that runs itself.

The hard part was
never the model.

The business model
was the AI's idea.

Nobody here
writes a word
or shoots a frame.

It is not a demo reel.
It grows every day.