Tabular Foundation Model · prediction without training

Tabular prediction, delivered with its proof.

Drop in a table, a CSV or a database extract, and Tabaxiom turns it into a predictive model in one second: imputation, scoring, anomaly detection. Every evaluation ships with its verifiable audit proof. No training, no data scientist, no GPU.

~1 stable → model served
CPUlocal, no GPU
0.779AUC · > gradient boosting
Technology

A foundation model for tabular data

Just as large language models learned text, a Tabular Foundation Model learned the structure of tables across millions of examples. It predicts on your columns without retraining, it reasons in context from your few rows.

01 / zero training

No training run to launch

The model reads your table and predicts. A new dataset means an immediate result, no hyperparameter tuning, no pipeline to maintain.

02 / small-N

Strong where data is scarce

Where classical methods demand tens of thousands of examples, it performs from a few hundred rows, the regime of real business cases.

03 / privacy

Local and on-premise

Runs on CPU, on your premises. The model saves without the training data, your data never leaves your infrastructure.

04 / granularity

One model per table

Every table, every client, every event gets its own instant model, without multiplying pipelines or operating costs.

05 / calibration

Calibrated and explainable

Calibrated probabilities, column importance, built-in quality control. Every prediction is auditable, not a black box.

06 / integration

Embeds anywhere

In your database (SQL function), your ETL, your CRM, your application, prediction where your teams already work.

Method, our difference

Every evaluation comes with its proof.

Anyone can plug in a tabular foundation model, they are public. Our difference: auditable trust, an evaluation protocol that cannot be gamed, inherited from the most demanding quantitative finance. In healthcare, insurance, credit, anywhere a prediction carries consequences, it is the line between “works on my laptop” and “I can deploy and defend it”.

protocol / sealed

Sealed before computation

The hypothesis is timestamped and hashed before any training. No rewriting history after the fact, no picking the convenient result.

protocol / time

Purged time

We test on strictly posterior data, with an embargo. The future never leaks into the past, the number-one trap of fake backtests.

protocol / canary

Anti-leak canary

We replay every test with shuffled labels. If the score still rises, there is a leak: the run is voided, not published.

protocol / verdict

Judged exactly once

One verdict, once, engraved. No “retry until it works”, the silent recipe behind most unverifiable results.

protocol / calibration

Calibrated probabilities

“30% risk” means 30%. You can size, provision, and decide on the number, not merely rank it.

protocol / baseline

Honest baseline

We always measure against the simple model to beat. No “+20%” out of thin air: the gain is real, quantified, reproducible.

Illustration, in your browser

Give it a table full of holes. Watch it fill them.

Drop in a CSV with empty cells, or run an example. This illustration runs 100% in your browser (nothing is sent) and gives a qualitative preview. It is not the Tabaxiom model: the real model runs server-side through the API, with its true accuracy and its evaluation proof.

Local illustration, not the full model. The real model runs server-side via the API.

Drag a CSV file here or

Predicted cells are highlighted. Live computation, no tricks.

Concrete examples

One capability, one input, one output

No abstraction: for each verb, here is the real call and what comes back. Same engine, same evaluation proof every time.

SELECT PREDICTDatabase · predict inside a SQL query
-- No extraction, no pipeline: prediction is a function. SELECT client_id, PREDICT(churn) AS risk, PREDICT_PROBA(churn) AS proba FROM clients ORDER BY proba DESC LIMIT 3;

→ Result

client_idriskproba
C-4471churn0,94
C-1180churn0,89
C-8023stays0,12
The probability is calibrated: “0.94” means 94% real risk, you prioritise follow-up on the number.
/imputeFill in a table's missing fields

Input, empty cells (∅)

citypopulationdensity
Bordeaux260 000
Lille6 612

→ Output, completed in context

citypopulationdensity
Bordeaux260 0004 629
Lille232 0006 612
Each value is inferred from the other columns of similar rows, not from a simple average.
/anomalySpot the aberrant in context (multi-column)

A trade that looks normal field by field…

instrumentqtypricevs mid
OAT 10Y5 M98,10+0,02
OAT 10Y5 M91,40−6,7
Each field is plausible on its own; the combination of size × price × distance to mid is not. Anomaly score 0.97, likely fat-finger.
/matchTwo records = same entity?
sourcenameid
InternalACME Corp.FR0000120271
BrokerAcme CorporationACME.PA

→ Verdict

same entityconfidenceproof
yes0,97a91f…3d
Learns from your past reconciliations, formats, ISIN vs ticker, casing, without hand-written rules.

These examples illustrate the target API contract (representative values); the signed proof/receipt field is on the roadmap. The only measured, sealed result on this site is the benchmark below, that is our honesty rule.

Proof, not promise

2,321 companies that truly delisted. A single verdict.

Hypothesis sealed before computation, time purged to forbid any leak, anti-cheat canary run, then measured. On real delisted companies, the foundation model beats gradient boosting, the industry yardstick.

Death-Oracle · purged test window · clean canary · judged once
ModelAUCTop-risk decile
Logistic regression (floor)0.7041.0×
Gradient boosting (yardstick)0.743
Foundation model (TabICLv2)0.7794.3×
What this verdict says

The foundation model achieves the best AUC and captures 4.3× more delistings in its highest-risk decile, the number that matters when prioritising follow-up.

Verdict sealed, judged exactly once, non-replayable. The reproduction is published in the internal claims register.

Honest scope: price/volume signal only, adding fundamentals is the next step. 34,000 training rows is mid-scale, not the “big data” regime where a well-tuned tree can pull ahead again. The verdict covers “foundation ≥ gradient”, not “finished product”. An audited proof also states its limits.

Use cases

From table to decision, across every industry.

The same technology, wherever there is a table and a target. 16 sectors, 83 documented cases. Expand a sector, then open the demo on a close example.

API, for your developers

A CSV goes in. Calibrated predictions, and their proof, come out.

Drop in a table, name the target column, a model served in tens of milliseconds, with its validity metric and its evaluation proof. No data science required on the calling side.

POST /v1/predict (target contract) { "table" : "clients.csv", "target" : "churn" } ← 200 OK · 41 ms { "pred" : 0.94, "calibrated" : true, "canary" : "clean", "receipt" : "a91f…3d" }
/predictpredict one column from the others, classification or regression.
/imputefill a table's empty fields, column by column.
/anomalyflag aberrant rows, including multi-column anomalies.
/matchdecide whether two records are the same entity, dedup, reconciliation.

Today: every evaluation includes its protocol, its baseline, its leak-test status and its diagnostics, which your compliance team can verify. On the roadmap: a signed, portable receipt on every API answer.

CPU · no GPUBatch & real timeOn-premise / GDPRPython SDKSQL function

Available today on-premise / locally. The hosted multi-tenant API is in early access, request your key.

Deployment

From your table to prediction in production

We identify the case
Which column to predict, which table, which business gain, scoping workshop.
We connect your data
CSV, SQL database, data warehouse: read-only connector, no migration.
Instant model
The foundation model serves your predictions, calibrated, audited, no training.
Integration
API, SQL function, or embedded in your tool. Local, cloud or on-premise.

What we deliver

An audited prediction service, honest baseline, anti-cheat canary, real metrics. Not a black box.

REST APISQL functionOn-premiseCPU / no GPUBatch & real timeGDPR / local
Frequently asked questions

What we get asked most

What is a tabular foundation model?

A model pre-trained on millions of tables that predicts on your columns without retraining: give it a few example rows and it answers in context. Just as a language model learned text, this one learned the structure of tabular data.

Do I need a GPU?

No. The model runs on CPU, locally or via the API. This site's illustration computes in your browser and is not the full model. No GPU to buy.

How is this different from XGBoost / AutoML?

Not raw accuracy, the deployment model: zero training, one model per table, strong on small data, cold-start, served on CPU. You can retune a tree; you cannot give it these properties.

Does it really beat gradient boosting?

On our sealed test (2,321 companies that truly delisted): AUC 0.779 vs 0.743, 4.3× in the top-risk decile. Hypothesis sealed before computation, judged once. See “Proof, not promise”.

Does my data stay private?

Yes. Local, on-premise or via the API; the model saves without your data. The site's public illustration transmits nothing and is not the full model.

How much data do I need?

A few hundred to a few tens of thousands of rows: the sweet spot. At millions of rows, a well-tuned tree can regain the edge, we say so.

What does “audited prediction” mean?

Hypothesis sealed before computation, purged time, anti-leak canary, judged once, calibrated probabilities. Every published evaluation comes with its verifiable proof; a signed receipt per API answer is on the roadmap.

Can we try it on our own data?

Yes: drop a CSV into the site demo for an instant preview, or contact us for an evaluation on your real data, measured honestly.

Evaluation on your data

You have a table. You have a case.

Tell us what you want to predict, we show you the result on your real data, measured honestly, with its evaluation proof.