Thnkrr — Video in Posts: How We Store It, How We Check It, What It Costs

What this is: the full record for video in posts (decision D-69, built 28 September 2026). It follows a video from the moment a student picks the file to the moment it plays in someone's feed: where it is kept, how it is checked, what is still to build, and what it costs. The one-page summary for the founder is 24-video-founder-summary.md.

Status: built and tested on the development branch. Not live yet. Prices checked on 29 September 2026 (AWS US East region, US dollars). Where a number is our own estimate rather than a price or a measurement, it says so.

In one minute

• A post can carry one video: up to 15 minutes and 5 GB, never together with photos. These are LinkedIn's own limits.
• Every video is rebuilt on our own server, so hidden location data from the phone can't survive. Its speech is turned into text, so a phone number said out loud is caught just like a written one.
• A person watches every video all the way through before anyone else can see it.
• Money: at the volume we expect, storing and delivering video costs about $5–20 a month, paid from our AWS credits. The automatic picture check we still have to add costs about $54 a month from Amazon, or nothing on our own server.
• The real cost is people's time: about 9–13 hours of reviewing a month at the volume we expect, and 33–50 hours in a busy month.
• Still to build: checks against lists of known child-abuse videos (free), detection of new child-abuse material (Thorn, price on request, or Google, free if approved), the automatic picture check, and some smaller fixes (section 7).
• Decisions we need from the founder: video length, which picture check, how long to keep originals, the free applications, and reviewer hours (section 7.3).

1. What a student can do

We copied LinkedIn's rules for member posts exactly (from LinkedIn's help pages, checked 28 September 2026).

The length and size limits are settings, not code. If the founder decides on a shorter limit for the pilot, it changes without a new release.

Posting a video, step by step:

1. Tap Add video in the post box and pick a file. The browser checks the limits it can read (type, size, length) before sending anything.
2. The file uploads straight away, with a progress bar. The student can keep writing.
3. "Processing your video…" — our server rebuilds it and drafts captions (section 2).
4. The student reviews it before posting:
   • a preview they can play;
   • a thumbnail: a frame we pick, or a picture they upload;
   • captions: the ones drafted from the speech, which they check and correct line by line; or a caption file they upload (the common ".srt" format); or none.
5. Post. The post waits for a person to check it (section 4).

What everyone else sees once it's approved:

• In the feed the video plays silently once most of it is on screen, one video at a time. Clips under 30 seconds loop.
• Captions are on by default, with buttons for captions, sound, and full size.
• Full size plays the sharper copy with sound, with a speed control (0.5× to 2×), and picks up where the feed left off.
• Anyone can turn off autoplay under Settings → Appearance → Autoplay videos. Devices set to "reduce motion" never autoplay.

Two deliberate differences from LinkedIn:

• No view counts. Thnkrr shows no popularity numbers anywhere (founder confirmed).
• Published posts can't be edited, as for every post on Thnkrr. A post the check rejected can be fixed: its text, captions and thumbnail can change, but never the video itself. To use a different video, delete the post and post again.

2. Where a video goes, step by step

Step 1 — Upload, into a locked room. The browser sends the file in pieces of 16 MB, three at a time, straight into private storage — a locked storage area only our server can open. We call it quarantine. If a piece fails, it is retried. If the connection drops, the upload carries on from the pieces already stored (as long as the same browser tab is still open). The upload goes straight into storage rather than through our server, so a big upload can't slow the site down.

Step 2 — Processing on our own server. Our server then fetches the file from private storage and, behind the scenes, one video at a time:

1. Reads what the file really is, and checks it against every limit in section 1 again. A file can lie about itself; the browser's word isn't trusted. A file that fails gets a plain explanation, like "Videos need to be at least 3 seconds long."

2. Rebuilds the video from scratch. It keeps exactly one picture track and one sound track and throws everything else away: location tags, phone make and model, chapters, subtitles, GPS tracks from action cameras. Then it checks the rebuilt file, and if anything other than a few harmless technical labels is left, the video is refused. (Tested on an iPhone-style file carrying a GPS location, "Apple" and "iPhone 17": all three were gone.)

3. Makes two copies for playing:

   • SD (480p, about DVD quality): what the feed plays. Small, quick to load.
   • HD (720p, basic high definition): what full size plays. Only made if the original is sharper than 480p. Nothing is ever enlarged.

   Both are ordinary MP4 files that play in every browser, at up to 30 frames a second.

4. Makes a poster (the still picture shown before playing) and a storyboard — 30 small frames taken evenly across the whole video, laid out in a grid. The storyboard is only for the person reviewing it.

5. Turns the speech into text, if the video has sound. This runs on our own server using Whisper: free speech-recognition software that anyone can download and inspect ("open-source"). AI software like this is called a "model". Because it runs on our server, no student's voice is sent to any other company. It only writes down the words. It never tries to work out who is speaking.

6. Saves the copies back into private storage and marks the video ready.

Step 3 — The student reviews and posts. Section 1, step 4.

Step 4 — A person checks it. Every video post goes to a human reviewer (section 4). The video's captions and the text of its speech are also read by the same automatic text check as every written post.

Step 5 — Approved: it goes public. The server copies the SD and HD files, the poster (or the student's own thumbnail) and a caption file into public storage. From there, Amazon's delivery network (CloudFront) serves them to viewers quickly. The copying happens inside storage; the video never passes through our main server.

Step 6 — Removed or deleted: it comes down. If a moderator removes the post, or the student deletes it, the public copies are deleted, not just hidden, so a saved link stops working too. The post is replaced by a short "removed" notice. The private copies stay, as the record of what happened.

3. What we store, and for how long

The SD and HD sizes were measured on a busy 3-minute 1080p test video (17.3 MB per minute for HD and 4.7 MB for SD). Ordinary footage usually comes out smaller. The "never more than" figures are the ceilings our settings allow.

4. How we keep video safe

How the reviewer checks a video:

• They first see the storyboard: 30 still pictures from across the whole video.
• The video stays blurred and silent until they choose to play it, so nothing shocking catches them off guard.
• The transcript and captions sit alongside.
• Approve stays locked until the whole video has played (up to 2× speed; skipping ahead doesn't count).

5. How reliable each check is

6. What it costs

6.1 Our assumptions

These are our own estimates, to be replaced with real numbers in the pilot's first month.

The viewing estimate assumes 1,500 people use Thnkrr each day. Each sees about four videos play silently for a few seconds, then opens one video full size and watches about half of it. For a busy month we assume about 2½ times as much watching per person (more videos, and longer ones).

6.2 Storage

Amazon storage (S3) costs $0.023 per GB per month.

Storage is cheap either way. Deleting originals after 30 days is still worth doing: it stops the one number that grows forever, and it means we hold less of children's data.

6.3 Delivery to viewers

Amazon's delivery network (CloudFront) is free for the first 1 TB every month, then $0.085 per GB.

The feed only downloads a video's poster until it starts playing, so videos scrolled past cost almost nothing.

6.4 Processing on our own server (measured)

We ran our real processing steps on a 3-minute 1080p test video with continuous speech, on a modern laptop processor (Apple M4). A cloud server's processor is slower; we assume 2–3 times slower, which is our estimate, not a measurement.

What that means for our server. The planned server is Amazon Lightsail's $12 plan: 2 processors and 2 GB of memory. It can run steadily at 20% of each processor, and go faster for short stretches using credit it saves up while quiet (Amazon's published figures). Over a month, that steady allowance is about 290 hours of computing.

The wait is about half the computing time because the server's two processors share the work. It assumes the server has saved-up credit; if that has run out, waits can be several times longer.

Cost: $0 extra at the volume we expect. The things to watch are bursts, not the monthly total. A run of long videos can use up the saved-up credit and slow down the whole site. If that happens, we move video processing onto a second $12 server (the same software, doing only background work). That costs $12 a month more, from the credits. (The $24 plan has the same processors and only adds memory, so it helps only if the server runs short of memory.)

6.5 Checking (moderation)

Reviewer time is the biggest cost of video. A reviewer plays each video in full (at 1–2× speed) and spends about a minute on the storyboard, transcript and decision.

A single 15-minute video takes a reviewer 8–16 minutes. The monthly totals depend on the average length, so a lower limit mostly protects against the worst cases: a burst of long videos, or someone uploading long videos on purpose to flood the queue.

The automatic checks:

Note on Transcribe: Amazon's own worked example uses $0.006 a minute; older price lists say $0.024. We haven't confirmed which applies to us, so the table shows both ends.

6.6 The total, per month

Everything from Amazon is paid from our ≈ $5,000 in AWS credits.

6.7 Running models on our own server, or paying a service?

We already run one model on our own server: speech-to-text. It costs no money, and uses about 1–2% of the server's steady allowance at the volume we expect. It also keeps every student's voice inside our own systems.

We could do the same for the automatic picture check with an open-source model:

• Money: $0 per video.
• Computing (our estimate, not yet measured): small models check a frame in about 0.01–0.03 seconds on a cloud processor; larger, more accurate ones take about 0.3–0.7 seconds. At one frame a second, that is between a couple of seconds and about 45 seconds of computing per minute of video — at most about 6 hours a month at the volume we expect, and 25 hours in a busy month. It fits on the same server.
• The catch is accuracy, not money. Our earlier research (docs/20 §10.3) found the open-source picture models check nudity only — no violence, weapons, drugs or hate symbols — and one popular model also labels faces, which clashes with our promise of no face analysis. Reading text written on screen would need a second open-source tool. And nobody stands behind their results but us.
• What can never run on our own server: child-abuse detection. No legitimate open-source model for it exists; those checks are only available from Thorn, Google, Microsoft and the child-protection bodies, under agreement.

Would running models ourselves raise costs a lot? No. There is no per-use charge, unlike a paid service. The one possible cost is the second server from 6.4: $12 a month more. What we give up is accuracy, and a supplier who is accountable for the results.

Our recommendation: use Amazon Rekognition for the picture check. It covers more kinds of harm, the credits pay for it, and it's the same work the photo check needs anyway. Build it so the supplier can be swapped without rework, as we did for speech-to-text, so an open-source model can replace it later if the cost or a privacy concern calls for it.

7. What's still to build

7.1 On the moderation side

7.2 Before going live

7.3 Decisions for the founder

Product decisions and paperwork only; technical choices are the team's (below).

1. Video length for the pilot. 15 minutes (LinkedIn's limit) or shorter. The cost is reviewer time (section 6.5); the setting changes without a release.
2. How long we keep the student's original file (the full-quality upload, locked away where nobody sees it): 30 days (recommended, except evidence in an abuse report) or forever.
3. Send the free child-abuse matching applications (7.1 #1) and ask Thorn for pilot pricing in the email already drafted (7.1 #2).
4. Reviewer hours: plan for about 18–25 minutes of video review a day at the volume we expect (counting every day of the month), and measure the real figure in week one.

Decided by the team: the picture check will be Amazon Rekognition, behind a connector an open-source model could replace later (section 6.7).

8. For engineers: where things live