Why Atelier?

Traditional actuarial pricing tools like Emblem are expensive, opaque, and tied to legacy platforms. Atelier is a modern, open-source alternative that wraps rustystats - a high-performance Rust-backed GLM engine - in a clean, interactive UI. It runs locally, stores everything on your machine, and follows the same explore-build-fit-iterate workflow actuaries already know.

Installation

uv add atel
# or
pip install atel

Installs everything - backend, frontend, engine. No separate build steps.

Quick start

atel                  # starts server, opens browser
atel --port 9000      # custom port
atel --no-browser     # start server only

The atelier command works too - atel is just shorter.

How it works

Workflow

Atelier follows the standard actuarial modelling workflow:

1. Upload - drag-and-drop a CSV or Parquet file, column types are auto-detected
2. Configure - select the response variable, GLM family, link function, offset, weights, and train/test split
3. Explore - pre-fit analysis runs automatically: response distribution, score tests ranking every candidate factor by expected deviance contribution, and a null (intercept-only) baseline model
4. Build - add terms from the factor sidebar: right-click any factor to choose categorical, linear, spline, target encoding, or other term types
5. Fit - hit fit, review the results: coefficient table, A/E charts, lift, calibration, VIF, and model diagnostics
6. Iterate - modify terms and re-fit. Every fit is auto-versioned so you can compare metrics across iterations and restore any previous version

Architecture

┌─────────────────────────────────────────────┐
│  Browser (React 19 + Tailwind + shadcn/ui)  │
└──────────────────┬──────────────────────────┘
                   │ HTTP/JSON
┌──────────────────▼──────────────────────────┐
│  FastAPI backend                             │
│  ├── /api/datasets   upload, validate        │
│  ├── /api/explore    EDA + null model        │
│  ├── /api/fit        GLM fitting             │
│  ├── /api/models     save, history, restore  │
│  └── /api/projects   project CRUD            │
├──────────────────────────────────────────────┤
│  rustystats          Rust GLM engine         │
├──────────────────────────────────────────────┤
│  SQLite (async)      projects, models, specs │
└──────────────────────────────────────────────┘

All data stays local at ~/.atelier/ - the database, uploaded datasets, and serialized models.

Features

Model building

• 8 GLM families - Gaussian, Poisson, Binomial, Gamma, Tweedie, Quasi-Poisson, Quasi-Binomial, Negative Binomial
• Rich term types - categorical, linear, B-splines, natural splines, target encoding, frequency encoding, expressions
• Monotonic constraints - enforce increasing/decreasing effects on splines and linear terms
• Interactions - standard product terms, target-encoded interactions, frequency-encoded interactions
• Regularization - Ridge, Lasso, Elastic Net with cross-validated alpha selection
• Train/test split - holdout validation with stratified splitting

Diagnostics

• Factor-level A/E - actual vs expected charts for every factor, fitted or not
• Score tests - chi-squared significance for candidate factors before fitting
• Lift charts - Gini, AUC, KS statistics with decile breakdown
• Calibration - Hosmer-Lemeshow test, decile calibration with confidence intervals
• Residual analysis - deviance, Pearson, and working residuals
• VIF & multicollinearity - variance inflation factors with severity coloring
• Model comparison - side-by-side train/test metrics against a base model

Data exploration

• Pre-fit analysis - response distribution, zero inflation, overdispersion detection
• Correlation matrix - numeric correlations and Cramer's V for categoricals
• Interaction detection - greedy residual-based search for potential interactions

Version control

• Auto-versioning - every fit is saved as a new version with full spec, coefficients, and diagnostics
• Change tracking - history panel shows terms added, removed, or modified between versions
• Restore - click any version to restore its terms and results, then continue iterating

License

Eclipse Public License 2.0