Predictive decision-making.Never been easier.

Tell it what you want to predict. It learns from the rows you have already judged, scores every remaining one, and puts a confidence figure on each. In chat, or as a tool your agents call.

See how it works Get in touch

~50: labeled rows to start
0: training runs
100%: on your hardware

pipeline_q3.csv — 4,182 rows local · CPU

you › predict which open deals will close this quarter

layer › found 52 labeled rows (Won/Lost). Scoring 4,130 remaining rows…

Account	Seats	Usage	Last touch	Prediction	Confidence
Northwind Freight	120	high	3d	labeled example · Won
Cobalt Health	18	—	41d	labeled example · Lost
Harbor Analytics	84	high	2d	0.91	94%
Meridian Labs	31	medium	9d	0.68	72%
Arcline Retail	—	low	27d	0.34	55%
Verity Logistics	210	high	1d	0.88	90%
Pinegrove Bank	12	low	63d	0.11	87%

Missing values kept as-is · no training run · results returned as CSV or over the tool API

Everything a prediction project usually needs — removed.

No data science team, no pipeline, no retraining schedule. Just the question and the rows you already have.

~50 examples is enough

Label a handful of rows the way you'd want them labeled. That's the whole setup — no feature engineering, no labeling marathon.

No training. No tuning.

There's no fit step, no hyperparameters, no model registry. Ask, and the remaining rows come back scored in seconds.

Confidence on every row

Each prediction carries its own confidence, so you can auto-act on the certain rows and route the rest to a human.

Plain CSVs, messy included

Point it at the export you already have. Blank cells, free-text columns and inconsistent types are handled, not rejected.

Runs on a laptop CPU

A tabular foundation model small enough to run locally. No GPU cluster, no queue, no per-prediction latency tax.

Data stays on your network

Your rows never have to leave your machine or your VPC. Compliance review becomes a conversation, not a project.

See it work

Three real prompts. Same model, different questions — browse with the arrows.

tracked_companies.csv — 4,012 rows local · CPU

● tracked_companies.csv 4,012 rows

I will treat those 84 owners as your examples of someone who sold, and score everyone else against them. Two things before I start. Does a sale to a strategic buyer count the same as a sale to a fund? And is there a size you stop caring below?

✓Read your list 4,012 companies, 84 marked sold

✓Kept the ones in range 3,180 with 20 staff or more

✓Found what the 84 had in common owner age, years held, flat headcount

✓Scored everyone else 3,096 companies in 7.4s

tracked_companies_scored.csvsorted by sell intent

company	owner age	years held	staff	state	sell intent	confidence
Northwind Mechanical	63	27	41	OH	high	0.92
Kestrel Heating	67	31	24	PA	high	0.81
Orchard Climate	59	22	63	NC	high	0.64
Vantage Air	45	12	130	FL	low	0.90
Aster Plumbing Co	38	6	88	TX	low	0.87

Download CSVYour original rows, untouched. Two columns added.

214 companies score above 0.80. That is your call list for the quarter. Ask me why any one of them scored the way it did, or tell me I got one wrong and I will take it into account.

names_only.csv — 1,120 rows · 2 columns local · CPU

● names_only.csv 1,120 rows · 2 columns

Yes, but I have to go and find the data first. For each name I will pull what is public: headcount, how long it has traded, hiring activity, reviews, licence records. Then I will label a sample myself, in writing, and you correct the ones I got wrong. That data is scraped, so it is right most of the time and not all of the time. How many companies do you want to end up ringing?

✓Matched your names to a public record 1,043 of 1,120

✓Enriched every one 14 fields per company, from filings, job posts and reviews

✓Labeled a sample itself 60 rows, each with a written reason

✓You changed 7, then it scored the rest 983 companies in 4.1s

enriched_and_scored.csvfive columns it found, two it predicted

company	town	staff	years trading	hiring	worth a call	confidence
Ridgeline Plumbing	Toledo OH	31	24	none	yes	0.74
Bellweather Heating	Erie PA	44	19	none	yes	0.68
Cedar Point Mechanical	Akron OH	26	31	none	yes	0.61
Meridian Air	Columbus OH	118	9	heavy	no	0.83
Talbot Services	Dayton OH	12	5	steady	no	0.79

Download CSVYour two columns, plus the fourteen it found and the two it predicted.

Your top 50 are ready. Three fields in ten came from a guess rather than a filing, so treat this as who to ring first, not as a fact about the business. Ask me where any number came from and I will show you the source.

past_deals.csv — 210 rows · 9 columns local · CPU

● past_deals.csv 210 rows · 9 columns ● current_pipeline.csv 12 rows

Your file has price paid and it has earnings, but there is no column for the multiple, and the multiple is the thing you are asking me to predict. I can work it out: price divided by earnings, on every row where both are filled. Two of your deals have no earnings and would drop out. Shall I derive it?

Done. 208 deals now carry a multiple, from 3.1x to 8.4x. That is thin for a number spread that wide. I can keep some of your deals back and price them blind, so you see how far off I am before you trust the twelve. How many should I hold?

✓Derived the multiple price ÷ earnings, on 208 of 210 deals

✓Held 40 back and priced them blind never saw what was paid

✓Checked itself against the real price off by 0.4x on average

✓Priced your current list 12 companies in 1.2s

current_pipeline_priced.csvsorted by multiple

company	revenue	ebitda	staff	years trading	multiple	likely range
Orchard Climate	$14.2m	$2.7m	63	22	6.1	5.2 to 7.3
Vantage Air	$26.8m	$4.1m	130	12	5.9	5.1 to 6.8
Northwind Mechanical	$9.4m	$1.6m	41	27	5.4	4.8 to 6.2
Aster Plumbing Co	$17.3m	$2.2m	88	6	4.9	4.0 to 6.0
Kestrel Heating	$5.1m	$0.8m	24	31	4.6	3.9 to 5.4

Download CSVA multiple and a range on every row in your pipeline.

Median 5.1x across your twelve. Two sit above anything you have paid before, which usually means I have no close comparison rather than that the business is special. On the forty deals I never saw, I was off by 0.4x.

Three steps, one sitting.

The model reads your examples in context. There is no fitting step to wait on and nothing to redeploy when the data changes — you just ask again.

Called as a tool

predict(
  file: "accounts.csv",
  target: "will_churn",
  examples: 52
)
→ 4,130 rows scored
→ confidence per row

01

Say what you want to predict

In plain language: “which of these accounts will churn next quarter?” No schema mapping, no target-column ceremony.
02

It learns from your labeled rows

The tabular foundation model reads the examples you already labeled — roughly 50 is plenty — and infers the pattern in-context.
03

Every remaining row comes back scored

A prediction and a confidence per row, in chat or returned straight to the agent that called it as a tool.

What teams point it at first

If the answer lives in a spreadsheet and a person could label 50 rows of it, it's a fit.

Deal & intent scoring

Rank open opportunities by likelihood to close from the CRM export you already run.

Lead & account scoring

Turn 50 hand-picked good-fit accounts into a scored list across the whole database.

Churn & renewal risk

Flag the renewals worth a call this week, with confidence to size the outreach.

Data-quality flags

Learn what a bad record looks like from a few examples and catch the rest automatically.

Category fill-in

Backfill missing categories, segments and tags across thousands of rows consistently.

Estimates & grades

Predict numeric estimates or graded tiers where a rule set would be too brittle to maintain.

Shipped by revenue, data and platform teams

“We had a churn model on the roadmap for two quarters. We labeled 60 accounts on a Tuesday and had scored renewals the same afternoon.”

Dana Whitfield

VP Revenue Operations, Harbor Analytics

“It's the first predictive tool our agents can just call. Structured scores with confidence beats an LLM guessing at a spreadsheet.”

Marcus Reyn

Head of Platform, Verity Logistics

“Nothing left our network, so security signed off in a week. That's never happened with an ML vendor before.”

Priya Nandan

Director of Data, Cobalt Health

Bring a CSV and 50 labeled rows. Leave with predictions.

Run it on your laptop, keep your data on your network, and find out in an afternoon whether the prediction is worth a project.

No GPU required · works offline · CSV in, CSV out