Norma treats preprocessing as a search-to-optimize problem. Every transform, encoding, and binning choice is a candidate pipeline. Each pipeline is scored by 5-fold XGBoost cross-validation. The best one wins.
Point it at a table. Name the target. Get back a model-ready dataset and the recipe that produced it.
When this matters
Norma is for teams whose data is messy, whose target is clear, and whose preprocessing is held together by convention.
Log the price, target-encode the country, bin the age. Your repos rhyme. The intuition is shared but never tested against a search.
XGBoost is a one-liner. Getting to the inputs that make it work is two weeks. The bottleneck moved a long time ago.
You'd try mean-encoding with smoothing, or splines on tenure, or a ratio feature. There was never time to A/B every variant against CV.
The training transform was eyeballed in a cell. The serving transform was reimplemented in Python. The scores never quite line up.
How it works
Norma frames preprocessing as optimisation. The objective is held-out CV score. Every choice is a coordinate in the search space.
A Parquet file, a warehouse query, a CSV. Norma reads the schema, samples the rows, and infers types.
One column. Classification or regression. Norma picks the right metric (AUC, RMSE, log-loss) and locks the folds.
It proposes candidate pipelines (encodings, scalers, bins, ratios, interactions) and scores each with 5-fold XGBoost CV. Bad branches die early.
You get the model-ready dataset and the recipe that built it. Same recipe runs at training and at serving. No drift.
What ships
Three artefacts come out of every search.
Parquet out, leakage-safe by construction. All target encodings are fit per fold; all imputers learn on the train side. Drop it straight into your trainer.
Every transform, every parameter, every seed. The recipe is the pipeline — load it at serving time and you get the same features for the same row.
log1p(price)
box-cox(tenure)
target_enc(country, smoothing=20, cv=5)
freq_enc(plan)
bin(age, 7, strategy=quantile)
price / (tenure + 1)
plan × country
Every trial scored, every CV fold logged, every feature attributed. You can read why the winner won and where the second-best fell off.
Why Norma
Two principles hold the system up.
01
Five-fold cross-validation, stratified, with target encodings fit inside the fold. If a feature wins, it wins on data the model never saw. No leakage, no story.
02
Reproducibility is a property of the artefact, not the notebook. The recipe Norma exports is what runs at serving — same code path, same parameters, same row-level features.
What Norma isn't
The space is crowded with adjacent tools. Here's what makes Norma structurally different.
Norma stops at the dataset. The model is yours — XGBoost, a linear baseline, a neural net, whatever. We optimise the inputs, you own the architecture.
A feature store serves features you've already defined. Norma generates new ones from raw columns and proves which ones move the metric.
No autocomplete, no in-cell suggestions. Norma runs as a non-interactive search — submit a job, get a leaderboard, ship the winner.
Common questions
XGBoost is a strong, fast, well-calibrated baseline on tabular data. Using it as the scorer means a feature that helps Norma is a feature that helps almost any tree-based or linear model downstream. The signal generalises.
Every transform that learns from the target (target encoding, mean encoding, smoothed counts) is fit inside the cross-validation fold, never on the full set. The recipe Norma exports preserves the same fold-aware logic at serving time.
The search runs on a representative sample by default — typically a few hundred thousand rows is enough to score pipelines reliably. The winning recipe then runs over the full dataset to produce the final artefact.
Yes. Whitelist or blacklist transforms, cap the number of generated features, fix specific encodings for compliance. The default search is broad; the constrained search is a flag.
Norma runs as a job — locally, in your VPC, or as a hosted service. The data never leaves the boundary you point it at. The output is a Parquet file and a serialisable recipe.
Pricing
The engine is open. You pay when you want it to run faster, share more, or sit inside your perimeter.
Plan 01
Free
Self-hosted, single-node
The Norma engine, the recipe runtime, and the CLI. Run it on your laptop or on a single box.
Plan 02
From $399/mo
For data and ML teams
Hosted search workers, shared leaderboards, and warehouse connectors. Run searches in parallel and share recipes across the team.
Plan 03
Custom
Single-tenant deployment
Self-hosted in your VPC, audit-grade lineage, SSO. For regulated environments and high-volume warehouses.
§ 03 · Engagement intake
01 / Start a brief
We staff small and senior, scope by phase, and end on a written deliverable. We don’t sell decks or hours.
If we’re not the right team for the job, we say so on the first call. The bar is production, not pitch.
02 / Where to find us
Montreal, Quebec
Satellite office