Auto Prediction

An agent labels your rows.
You just check its work.

Labeling is where prediction projects die. Not the model, the afternoon nobody has for judging four hundred rows by hand. Auto Prediction puts an agent on that job, makes it show its reasoning, and lets you overrule it.

run  /  who-is-worth-a-call finished in 4m 12s
01 Collected the data 11 signals · 3 sources
headcount + 12mo change open roles by function funding stage tech stack review volume site traffic band office count

You gave it 4,120 company names. It worked out which signals would separate a good call from a bad one, went and got them, and wrote 38,400 fields back to your rows. Every field says where it came from.

02 Labeled a sample 200 rows
sample_labeled.csv200 of 4,120
companylabelwhy it said so
Northwind Ltdworth a call Hiring three ops roles, headcount up 40% in a year
Aster Healthnot yet Eight people, no ops function to sell into
Kestrel Roboticsworth a call Two warehouses opened since March, stack matches
Vantage Printnot yet Flat headcount, no hiring, legacy stack
6 labels you corrected, folded back into the criterion

It labels against your criterion and writes down its reason each time. You skim, fix what it got wrong, and your fixes become the definition. Ten minutes of your attention instead of an afternoon.

03 Scored the rest 3,920 rows
companies_scored.csvDownload CSV
companyheadcountopen rolesstage predictionconfidence
Halcyon Group3107Series B worth a call0.94
Bellweather Co440Seed not yet0.89
Orchard Labs1,18021Series D worth a call0.86
Merrow Freight962Bootstrapped worth a call0.58

Every remaining row scored, with a confidence figure on each. Sort by it and you know which rows to trust and which to read yourself.

The part that was never automated

Labeling, done by an agent that shows its reasoning

Every prediction tool on the market starts by asking you for labeled examples. Nobody asks where those come from. They come from a person reading rows one at a time, which is why most prediction projects stall before the model is ever run.

An agent can read a row the way you would. Not just the numbers in the columns, but the filings, the job posts, the site. So we put it on the labeling job, and made it write down its reason for every single row. That reason is what makes the work checkable in minutes rather than hours.

Agent labeling  /  is this company worth a call 200 rows sampled from 4,120
Northwind Ltd worth a call Hiring three ops roles, headcount up 40% in a year agree
Aster Health not yet Eight people, no ops function to sell into agree
Kestrel Robotics worth a callnot yet You said: two new warehouses beats a legacy stack you fixed this
Vantage Print not yet Flat headcount, no hiring, legacy stack agree
200
rows labeled by the agent, each with a written reason
6
you disagreed with, which is all the work you actually do
10 min
to check the lot, instead of an afternoon judging rows cold

Why the reasons matter

A label you cannot check is worthless

An agent that just hands back four hundred labels has moved the problem, not solved it. You would have to trust it blind or redo the work. The reason column is what turns it into something you can actually accept.

It reads more than the columns

Filings, job posts, the company site, whatever it collected in step one. It labels on the same evidence a person would use, not on four numbers in a spreadsheet.

One line of reasoning per row

Plain English, short enough to scan. Wrong reasoning is obvious at a glance even when the label happens to be right, which is the failure everyone else hides.

Your correction is final

Overrule it and that row is settled for good. It relabels the rest of the sample against what you just told it, so one fix moves everything like it.

The gap

The model was never the hard part

Ask anyone who has shipped a prediction where the months went. Almost none of it was the model. It went on finding data that was not in the table, arguing about what the label actually means, and hand-tagging rows until somebody lost patience and shipped a rule instead.

Auto Prediction takes all three jobs. You say what you want to know. It works out what it needs, goes and gets it, labels a sample, checks those labels with you, and scores everything else.

  • Find the dataA list of names is not a dataset. It works out which signals would separate your outcomes and goes and collects them.
  • Agree the labelYour criterion lives in your head. It writes a first draft, you correct it, and the corrected version is what gets used.
  • Score the restA tabular foundation model reads your examples as context and scores every other row. Seconds, not a training run.

How it works

Say it once. Watch it work.

01

It finds the data

You have names and not much else. It decides which signals matter for your question and collects them from public pages, filings, job posts, review sites, and your own systems. Every field it adds carries its source.

02

It labels the examples

It reads a sample and labels each row against your criterion, with a one-line reason. You skim, correct what it got wrong, and your corrections outrank its guesses from then on.

03

It predicts the rest

Every remaining row scored, with a confidence figure on each. Download it, push it back to your CRM, or let your agent read it straight from the tool.

Control

Nothing happens off-screen

An automated pipeline you cannot inspect is a rumour with a progress bar. Every stage of a run is written down and open to you.

Change your mind about the criterion and run it again. It costs minutes, so you can afford to be picky about what a good row actually looks like.

Every field shows its source

Click any collected value and see the page, filing, or system it came from, and when it was read.

Every label shows its reason

One line per row, in plain English. If the reason is wrong, the label is wrong, and you can see it in a glance.

Your corrections win

A row you have judged is never overwritten by a row it guessed. Your edits define the criterion.

Where you ask

Ask from wherever you work

In chat

Describe the prediction in a sentence. It comes back with the run, the labels, and the scored rows.

From your agent

Over MCP, as a tool your agent calls mid-task. It gets a number and a confidence back, not a paragraph of prose.

In the browser

Watch a run as it happens, correct the labels by hand, and download the result.

Tell us what you want to know

Bring a question and a list of rows. We will show you what a run looks like on your own data.