KNOWWYour current approach is on the board
Every run scores a baseline, a plain regression, gradient-boosted trees and foundation models on the same time-based folds. The simpler model wins when scores are within one standard error.
Predictions from the tables you already keep.
Upload a spreadsheet of customers, loans, machines or orders. The Workbench runs today's tabular foundation models and gradient-boosted trees against your own history, validates them the way your reviewers expect, and shows which one to trust.
Free for a month: five runs, up to 50,000 rows each. No card, no installation.
| loan_id | opened | months | utilisation · 6m | late · 12m | fell_behind_90d |
|---|---|---|---|---|---|
| L-1041 | 2024-03 | 31 | 3 | yes | |
| L-1178 | 2024-07 | 27 | 0 | no | |
| L-1302 | 2025-01 | 21 | 1 | no | |
| L-1455 | 2025-06 | 16 | 2 | yes | |
| L-1590 | 2025-11 | 11 | 0 | blankto predict | |
| L-1627 | 2026-02 | 8 | 1 | blankto predict |
See a run
Two runs, start to finish.
Two seeded projects, recorded end to end. A table like yours runs the same way.
After one run
What you leave with.
Every run produces the same four things, on your own table, inside your own workspace.
A score and a range for every row
Yes/no questions get a probability with a calibration check. Numbers get an 80% range. Each row shows the strongest drivers behind it.
A threshold you can reason about
Move the cut-off and see approvals, expected misses and the business impact on your own base rate.
Watched from day one
The chosen model is frozen at a deployment date and replayed forward. Drift in any input raises a warning or an alert.
How a run works
One table in. Predictions out.
No feature pipeline to build and no model to configure. Bring the table, confirm three guesses, press Run.
- Bring one table
One row for each thing you want a prediction for. The outcome you want to know as a column, blank where it is not known yet; those rows get predicted. An ID per row and a date if you have one, so every test trains on the past and scores the future.
- Confirm three guesses
The Workbench proposes the target, the ID and the date column. Check them and press Run. Columns that look like personal data are flagged so you can drop them before anything is stored.
- Read the leaderboard, download the predictions
Most runs finish in a few minutes. Every decision the feature screen made is listed, with the with-and-without comparison.
The questionWill this account default within 90 days?
| ID | Months on book | Utilisation · 12m | Payments · 12m | Income Δ | Probability · 90% band | Status |
|---|---|---|---|---|---|---|
| 38 | Utilisation 32%, 35%, 38%, 41%, 47%, 52%, 55%, 61%, 66%, 72%, 77%, 81% | -18% | 0.42 | High | ||
| 64 | Utilisation 21%, 20%, 22%, 19%, 20%, 21%, 20%, 18%, 20%, 19%, 21%, 20% | +4% | <0.01 | Low | ||
| 12 | Utilisation 48%, 62%, 41%, 66%, 52%, 58%, 44%, 69%, 51%, 60%, 57%, 55% | 0% | 0.02 | Low | ||
| 1 | Utilisation 40% | — | 0.18 | Review | ||
| 27 | Utilisation 70%, 68%, 66%, 62%, 60%, 57%, 53%, 50%, 48%, 45%, 42%, 40% | +9% | <0.01 | Low | ||
| 45 | Utilisation 60%, 62%, 65%, 68%, 70%, 74%, 77%, 80%, 83%, 86%, 88%, 90% | -6% | 0.48 | High | ||
| 19 | Utilisation 30%, 31%, 30%, 29%, 30%, 31%, 30%, 30%, 29%, 30%, 31%, 30% | +2% | 0.01 | Low | ||
| 8 | Utilisation 10%, 12%, 11%, 10%, 14%, 38%, 55%, 64% | -30% | 0.13 | Watch |
On timeLateMissedNo history
accounts scored8
Flagged for review1
Leaderboard winnerFoundation model
Forward passes1
An illustration of what one scored row looks like: the probability, its range, the drivers behind it and a status. Switch the table to see the same format on a different question. All rows and scores here are synthetic examples; the real thing, on your own data, is in the Workbench.
Built for review
Evidence your reviewers already ask for.
Whether the reviewer is a bank's model-risk team or a COO, a run shows its work. We start with financial services, the strictest review there is.
Backtests
Validated on time, not on luck
Folds train on the past and score the future. A leakage test scrambles every row after the cut-off and checks that no feature changes.
Baseline
The baseline is always on the board
The naive answer, a plain regression and gradient-boosted trees sit next to the foundation models. Sometimes the simpler one wins. Then it wins.
Uncertainty
Ranges, calibration, drivers
80% ranges on numbers. Calibration curves on probabilities. The strongest drivers for each row, signed.
Monitoring
Watched after deployment
Population stability on every input, with warn and alert thresholds, and the score replayed month by month.
How every pilot runs
- Shadow mode
The model scores live cases. Your current process keeps deciding.
- Side by side
Both compared on the same cases.
- Validation sign-off
Your reviewers check backtests, calibration and drivers.
- LIVE
In production, with your current process as the fallback for rows flagged for review.
Results from the free month are for evaluation. Production use needs the review your field requires, such as a fair-lending or model-risk review, and that is what the pilot is for.
Foundation models that read your rows as context.
Prior-data fitted networks are trained once across millions of synthetic datasets. Shown a new table, they find the pattern in a single forward pass, with no training on your data and no model to maintain.
01Train · once
synthetic datasets × millionstransformer
A transformer learns from millions of synthetic records: what trends, seasonality, noise and shocks look like.
02Infer · one forward pass
01 / 09 · Rising trendFinds the pattern, shows its uncertainty. No retraining.
A published result on tablesHollmann et al., Nature 637 (2025)
Prior-data fitted network, out of the box
2.8s
Experts tuning models
4h
Hollmann and colleagues showed a prior-data fitted network matching four hours of expert tuning in seconds. Models of that family are what the Workbench runs: three tabular foundation models and one time-series foundation model, next to gradient-boosted trees, with the leaderboard deciding. How it works
Who it's for
Anyone with a table and a question about tomorrow.
Every industry runs on rows and columns that change over time. The same run answers each one's question. We start with financial services.
Lending · firstWhich clients will miss next month's payment?Default · collections · small-business loansMarketing & retentionWhich customers will take up an offer if we call?Cross-sell · churn · demandMaintenanceWhat will each machine cost next month?Cost forecasts · failure risk · facility totalEnergy & operationsHow much will the plant use tomorrow, hour by hour?Load · consumption · yieldInsuranceWhich claims need a second look?Claims · renewals · pricingSupply chainWhich shipments will arrive late?Delays · supplier risk · lead timesYour tableWhat do you need to predict?One row per thing, one outcome column. Tell us the question →
Your data
Nothing leaves our server.
Every model runs on our own machine. No row of your data goes to any outside service.
- Upload
Columns that look like personal data are flagged at upload, and you can drop them before anything is stored. The models do not need names, emails or ID numbers.
- Access
Only your workspace sees your projects. The sample projects are the one thing every workspace shares, and they use public or generated data.
- Retention
Files are kept for 30 days after the last run, or deleted the moment you press Delete. HTTPS on the way in, an encrypted disk on the server.
- Pilots
A pilot can run inside your own environment, with the same leaderboard, ranges and monitoring. Ask on the form.
Start with a month on us
Five runs on your own tables. Free.
Everything the Workbench produces, on a table you choose, with nothing to install. We ask for feedback in return.
5runs in the month
50,000rows per upload
500columns per upload
30days in your workspace
Included
- Any CSV or Excel table
- The leaderboard, with your current approach on it
- A predictions file to download, with a range and drivers per row
- The monitoring page and the what-if panel
- No card, no installation
After the month
When a run shows something worth acting on, the next step is a pilot next to your current model, in our environment or yours. Tell us what you need and we reply with a plan and a price.
Request your free monthQuestions
What teams ask first.
Short answers to the questions that come up before the first upload. The guide inside the Workbench covers every screen.
Something else?hello@knoww.ai
One table, one row per thing, the outcome as a column, blank where it is not known yet. An ID for each row, and a date if you have one. CSV or Excel, up to 50,000 rows and 500 columns in the free month.
Where this is going
The table is the next frontier.
Today's tabular models read a snapshot. Knoww is building the one that reads a history.
In 2026 SAP bought the lab behind TabPFN and Nvidia bought the company behind the first relational foundation model. Both bets say the same thing: the valuable prediction problems in a business live in tables, and a model can be pretrained to learn them from the rows you hand it.
LTFM, the model Knoww is building, is a longitudinal tabular foundation model. It reads entities, their state and their history, so the same model answers questions about change. The Workbench is where that work meets real tables.
Request your free month
What do you want to predict?
Five runs, 50,000 rows each, free for a month. Tell us the question and what you use today. We review every request by hand and email you when your workspace is ready. If you already know you want a pilot next to your current model, say so below.