Skein by RyanAI
Beta review · September 2026
getskein.io

Would you mind trying to break this?

Fifteen minutes, nothing to install. You are getting the running application, not a slide deck about one. I would rather you find the problems than a customer does, so please be unkind about it.

The application The real thing, running. Sign-in takes about twenty seconds, steps below. app.getskein.io No sign-in needed A million records, searchable in your browser. Good for two minutes on a phone. getskein.io/demo-large

Getting in the first time

One-time code, no password to invent
  1. Open app.getskein.io. A sign-in page appears before the application loads. That page is the door, not the product.
  2. Enter the email address this was sent to. Any other address is refused on purpose.
  3. A six-digit code arrives in a few seconds. Paste it.
  4. You are in, and stay in on that device for a day. No account is created and no password exists.

If it tells you that you do not have access, your address simply is not on the list yet. Reply and it takes a minute to add.

What it is, in one paragraph

Skein reads the records a business already has, meaning email, documents, spreadsheets and their live database, and writes the page you would have written before a client call. What you charge them and what that replaced, the figures on record, what is open and late, who is involved, and a section saying plainly what it does not know. It runs on the customer's own machine, calls no AI service, and sends nothing anywhere.

The fifteen minutes

  1. Pick a business from the list and read its briefIs this the page you would want in front of you before a call? More to the point, what is missing from it?
  2. Read the panel headed "what this does not know"Does that build trust with you, or does it read as an excuse?
  3. Search for something specificA person, a business, a phrase from a document. The timing shown is measured on your machine, not quoted from ours.
  4. Open the graph and click through itSolid lines were stated in the source data. Dashed ones we inferred. Is that distinction useful, or is it noise on the screen?
  5. Switch tenants using the selector at the topEach one is a separate business with its own data. Nothing should ever cross between them. If you can make something cross, that is the most valuable bug you could hand me.
  6. Then open the million-record demoSame interface, 1,009,300 records, running entirely in your browser tab. Does it still feel instant?

Two things are deliberately switched off on the hosted copy: importing data and reaching out to a database. It is sealed so that a public address cannot be pointed at anything. Run the download if you want to see the import, and it will connect to a real database in about four minutes.

The largest thing we have put through it

A live PostgreSQL holding 15,699,911 rows across 2,124 MB, indexed on one laptop in 2.2 minutes into a graph of 900,000 nodes. The part worth pressing on: fifteen million of those rows cost three seconds, because invoices and events are summarised per customer inside the database and never travel. The import is paid for by the 700,000 rows that become entities and by nothing else.

15,699,911rows indexed on one laptop
2.2 minthe whole first import, 278 MB peak
0.12 sre-index when nothing changed
194 MBpeak memory, flat from 250k to 4M rows
63 msmedian search on the finished graph
0.4 msto assemble one brief

How it stays current, which matters more than how fast it loaded

A one-off import is a snapshot: right on the day it runs and wrong every day after. We keep a watermark of how far the last run got, on a column that reveals edits rather than only new rows, and start the next run there. That is why an hourly cadence is affordable, and it is the thing I would most like you to try to break.

Hourlythe default cadence for a live source
0.12 swhat an hourly run costs when nothing changed
Edits toonot just new rows, and we say which kind a source gives us
2.7 sour own graph reading 1,367 documents off a second machine

We run this on ourselves. Our own graph pulls from a second machine on an hourly timer: a full read is 2.7 seconds, and a run where nothing changed is 0.3 seconds and opens no files at all.

Every one of those is reproducible from a command in the repository. Read the full report.

The questions I actually want answered

Where did it feel unprofessional or half-finished?

What would stop you paying $9,500 for this?

Is the brief genuinely the thing you would want before a call, or is it a dashboard wearing a document's clothes?

Did anything make you distrust the numbers?

Everything else

The siteWhat it is, how it works, price ExplainerPlain language, prints to PDF Performance report15.7M rows, per table, with the defects Claims, and how to break themEach one with the command that would disprove it The deckEight slides, for a room Run it yourself1.4 MB. Install Python, unzip, double click. Mac or Windows

What we do not have. No SOC 2 and no HIPAA agreement. For an install where the data never leaves the customer's own machine the audit surface is theirs rather than ours, which is an argument and not a certificate.

Matching up names is genuinely hard, everyone in this field claims to have solved it, and nobody has. Skein puts the doubtful ones in a queue and asks. Measured more accurate than deciding automatically.

One person has run the onboarding, once. That is the part I would most like you to try cold, without asking me anything, and tell me where you stopped.

Every name and figure in the demo is synthetic: no customer data appears anywhere. Skein by RyanAI LLC · Missouri · ry@ryanai.io