Smart-if — The decision factory

A small decision doesn't need a giant.

Large AI models can do a bit of everything, for everyone, from distant servers, and they bill every question. But your company mostly makes the same decisions, thousands of times a day.

For those, you need a specialist: a model chosen and trained for your task, that knows your categories and your vocabulary, runs on your infrastructure and answers in a few tens of milliseconds. And whose reliability you know, because we measure it before we deliver.

The method

Step 1
Measure the task
Which decision, what volume, which languages, which constraints? We compare several approaches on a sample of your data, including the simplest ones.
Step 2
Label with you
Your history is often a training set already. A model pre-labels, your teams validate the doubtful cases, without weeks of data entry.
Step 3
Train and compare
We train the right model, calibrate it and compare it with the best alternative. You get a quantified report, limitations included.
Step 4
Deliver on your premises
A container ready to install on your infrastructure, CPU or GPU. No data leaves, no cost per question.

Services

A three-stage path: start small, and commit only once the measures are in.

Diagnostic

Analysis of the decision and the data, measured comparison of several approaches on your sample, written recommendation. Including “you don't need a model” when that is the answer.

Quote on request, two to three weeks

Build

Assisted labelling, training, quality report, delivered container and support for integration into your systems.

Quote on request

Operate

Quality monitoring in production, drift detection, retraining when your data evolves, support.

Subscription, quote on request

Options

Compact model (sal-torch)

A training method that makes the model robust to compression: smaller and faster at equal quality, for more modest servers.

Certification (sal-certify)

A certification report for the delivered model, for your control and compliance teams. Detailed content on request.

Measure the task. Not the hype.

Our figures come from public datasets, with a published protocol. Your results will be measured on your own data, before any decision.

Accuracy on public test sets (public data, CC BY 4.0 licence). Specialised model: the model actually served on this site, measured by us. Other columns: decisionfactory evaluation report, October 2026.
Task	Specialised model	Classic method	API reference
Banking intents, 77 categories (English) · 500 sentences	0.924	0.894	0.798
Routing, 60 categories · 2,974 sentences
French	0.878	0.839	0.787
German	0.863	0.829	0.790
Italian	0.876	0.839	0.795
English	0.893	0.846	0.811
Yes / no question: is it about the calendar? · 500 sentences
French	0.964	0.908	0.930
German	0.968	0.932	0.914
Italian	0.968	0.932	0.914
Median latency per decision	~35 ms	2 to 56 ms	~220 ms network included
Cost per million decisions	~$0.09	< $0.02	$15 to $80

Latencies of the specialised model and the classic method: measured on our 4-vCPU server without GPU, container limited to 2 CPUs and 2.5 GB. API reference latency: measured client-side, network included, October 2026. Cost of the specialised model: full server price ($6.00 per month) attributed to the playground alone, server at 100 % load; a conservative estimate. Comparison costs: public prices as of October 15, 2026. The classic method remains the best option when a few points of accuracy matter less than cost: that is what the diagnostic is for. Read the full protocol

Your decisions stay with you.

The model runs on your infrastructure, or with a Swiss host if you prefer. It depends on no external API, always gives the same answer to the same question, and every delivered version is documented.

Tested in French, German, Italian and English on public data
Delivered in an open format, with no vendor lock-in
Quality and calibration report with every model

Specialised models for measurable decisions.

A small decision doesn't need a giant.

The method

Measure the task

Label with you

Train and compare

Deliver on your premises