millwright 0.1.1

A unified ML framework for Rust — proven Rust crates, assembled into one machine.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
<html><head><meta http-equiv="Content-Type" content="text/html; charset=UTF-8"><title>Millwright · Guide</title>
<meta name="viewport" content="width=device-width, initial-scale=1">
<style>
  :root{
    --bg:#EEF1F3; --surface:#FFFFFF; --surface-2:#F5F8F9; --surface-3:#E9EEF1;
    --ink:#16191C; --ink-soft:#556069; --ink-faint:#8894A0;
    --line:#DCE3E8; --line-soft:#E7ECEF;
    --oxide:#BC4A1E; --oxide-tint:#F6E6DE; --oxide-ink:#94380F;
    --patina:#2C8272; --patina-tint:#DCEEEA;
    --steel:#3E5566;
    --code-bg:#1A2024; --code-ink:#D7E0E6; --code-key:#E4703F; --code-str:#7CB8A6; --code-com:#6E7C86; --code-fn:#9DB8CC;
    --font-sans:system-ui,-apple-system,"Segoe UI",Roboto,Helvetica,Arial,sans-serif;
    --font-mono:ui-monospace,"JetBrains Mono","Cascadia Code","SF Mono",Menlo,Consolas,monospace;
    --measure:68ch;
    --radius:10px;
    --shadow:0 1px 2px rgba(20,30,40,.05),0 6px 20px -12px rgba(20,30,40,.18);
  }
  :root:not([data-theme="light"]){
    @media (prefers-color-scheme:dark){
      --bg:#13171A; --surface:#1A2024; --surface-2:#20272C; --surface-3:#161C20;
      --ink:#E7ECEF; --ink-soft:#9DA8B0; --ink-faint:#6C7882;
      --line:#2A333A; --line-soft:#232B31;
      --oxide:#E4703F; --oxide-tint:#2A1C15; --oxide-ink:#F0906A;
      --patina:#54B7A1; --patina-tint:#132723;
      --steel:#87A2B6;
      --code-bg:#10151A; --shadow:0 1px 2px rgba(0,0,0,.3),0 10px 30px -16px rgba(0,0,0,.6);
    }
  }
  :root[data-theme="dark"]{
    --bg:#13171A; --surface:#1A2024; --surface-2:#20272C; --surface-3:#161C20;
    --ink:#E7ECEF; --ink-soft:#9DA8B0; --ink-faint:#6C7882;
    --line:#2A333A; --line-soft:#232B31;
    --oxide:#E4703F; --oxide-tint:#2A1C15; --oxide-ink:#F0906A;
    --patina:#54B7A1; --patina-tint:#132723;
    --steel:#87A2B6;
    --code-bg:#10151A; --shadow:0 1px 2px rgba(0,0,0,.3),0 10px 30px -16px rgba(0,0,0,.6);
  }

  *{box-sizing:border-box}
  html{-webkit-text-size-adjust:100%; scroll-behavior:smooth}
  body{
    margin:0; background:var(--bg); color:var(--ink);
    font-family:var(--font-sans); font-size:17px; line-height:1.65;
    -webkit-font-smoothing:antialiased; text-rendering:optimizeLegibility;
  }
  .wrap{max-width:1080px; margin:0 auto; padding:0 24px}
  .col{max-width:var(--measure)}

  h1,h2,h3{line-height:1.15; text-wrap:balance; margin:0; letter-spacing:-.012em}
  h2{font-size:clamp(1.5rem,1.1rem + 1.5vw,2rem); font-weight:650; margin-bottom:.5rem}
  h3{font-size:1.12rem; font-weight:650; margin:1.4rem 0 .3rem}
  p{margin:0 0 1rem; max-width:var(--measure)}
  a{color:var(--oxide-ink); text-underline-offset:2px; text-decoration-thickness:1px}
  strong{font-weight:650}
  .eyebrow{
    font-family:var(--font-mono); font-size:.72rem; font-weight:600;
    letter-spacing:.14em; text-transform:uppercase; color:var(--oxide);
    display:inline-flex; align-items:center; gap:.5rem;
  }
  .eyebrow::before{content:""; width:26px; height:1px; background:var(--oxide); opacity:.6}
  .mono{font-family:var(--font-mono)}
  .muted{color:var(--ink-soft)}
  .tiny{font-size:.86rem; color:var(--ink-soft)}

  header.top{
    position:sticky; top:0; z-index:20; background:color-mix(in srgb,var(--bg) 86%,transparent);
    -webkit-backdrop-filter:saturate(1.4) blur(8px); backdrop-filter:saturate(1.4) blur(8px);
    border-bottom:1px solid var(--line);
  }
  .top .wrap{display:flex; align-items:center; justify-content:space-between; height:56px}
  .brand{display:flex; align-items:baseline; gap:.6rem; font-family:var(--font-mono); font-weight:700; letter-spacing:-.02em}
  .brand a{color:inherit; text-decoration:none}
  .brand .mark{color:var(--oxide); font-size:1.15rem}
  .brand .ver{font-size:.72rem; color:var(--ink-faint); font-weight:500; letter-spacing:.05em}
  .top nav{display:flex; gap:.95rem; font-family:var(--font-mono); font-size:.74rem; letter-spacing:.03em}
  .top nav a{color:var(--ink-soft); text-decoration:none}
  .top nav a:hover{color:var(--oxide-ink)}
  .top nav a.repo{color:var(--oxide); font-weight:600}
  @media(max-width:680px){.top nav{display:none}}

  .hero{padding:72px 0 32px}
  .hero h1{font-size:clamp(2.3rem,1.6rem + 3vw,3.6rem); font-weight:700; letter-spacing:-.03em}
  .hero .sub{margin-top:.35rem; font-family:var(--font-mono); font-size:.86rem; letter-spacing:.02em; color:var(--patina)}
  .hero .lede{margin-top:1.4rem; max-width:62ch; font-size:1.22rem; color:var(--ink); line-height:1.5}
  .hero .lede b{color:var(--oxide-ink); font-weight:650}

  section{padding:38px 0; border-top:1px solid var(--line-soft)}
  section > .wrap > .head{margin-bottom:1.3rem}
  .head .col{max-width:62ch}

  pre{
    background:var(--code-bg); color:var(--code-ink); border-radius:var(--radius);
    padding:18px 20px; overflow-x:auto; font-family:var(--font-mono); font-size:.84rem; line-height:1.7;
    border:1px solid color-mix(in srgb,var(--code-ink) 12%,transparent); margin:0 0 1rem;
  }
  pre .k{color:var(--code-key)} pre .s{color:var(--code-str)} pre .c{color:var(--code-com)} pre .f{color:var(--code-fn)} pre .d{color:var(--ink-faint)}
  code.inl{font-family:var(--font-mono); font-size:.88em; background:var(--surface-3); padding:.08em .38em; border-radius:5px; color:var(--oxide-ink)}

  .tablewrap{overflow-x:auto; border:1px solid var(--line); border-radius:var(--radius); margin-bottom:1rem}
  table{border-collapse:collapse; width:100%; font-size:.9rem; min-width:520px}
  th,td{text-align:left; padding:10px 14px; border-bottom:1px solid var(--line-soft); vertical-align:top}
  thead th{font-family:var(--font-mono); font-size:.7rem; text-transform:uppercase; letter-spacing:.08em; color:var(--ink-faint); font-weight:600; background:var(--surface-2)}
  tbody tr:last-child td{border-bottom:0}
  td .feat{font-family:var(--font-mono); color:var(--oxide); font-size:.86rem}
  tr.core td{background:var(--oxide-tint)}
  .badge{font-family:var(--font-mono); font-size:.62rem; letter-spacing:.06em; text-transform:uppercase; padding:.12rem .4rem; border-radius:5px; background:var(--patina-tint); color:var(--patina); margin-left:.4rem; vertical-align:middle}

  ul.clean{list-style:none; margin:.2rem 0 1rem; padding:0; display:flex; flex-direction:column; gap:.5rem; max-width:var(--measure)}
  ul.clean li{position:relative; padding-left:1.2rem; font-size:.95rem; color:var(--ink-soft); line-height:1.5}
  ul.clean li::before{content:""; position:absolute; left:0; top:.62em; width:6px; height:6px; border-radius:1px; background:var(--oxide)}
  ul.clean li b{color:var(--ink); font-weight:600}

  /* ---- guide-specific ---- */
  .toc{display:grid; grid-template-columns:repeat(3,1fr); gap:8px; margin-top:.6rem}
  @media(max-width:760px){.toc{grid-template-columns:repeat(2,1fr)}}
  @media(max-width:460px){.toc{grid-template-columns:1fr}}
  .toc a{
    display:flex; gap:.55rem; align-items:baseline; text-decoration:none; color:var(--ink);
    background:var(--surface); border:1px solid var(--line); border-radius:8px; padding:9px 12px;
    font-size:.9rem; box-shadow:var(--shadow);
  }
  .toc a:hover{border-color:var(--oxide)}
  .toc a .num{font-family:var(--font-mono); font-size:.72rem; color:var(--oxide); font-weight:600}
  .eyebrow.n::before{width:16px}
  .callout{
    background:var(--surface-2); border:1px solid var(--line); border-left:3px solid var(--patina);
    border-radius:8px; padding:12px 15px; margin:0 0 1rem; font-size:.92rem; color:var(--ink-soft); max-width:var(--measure);
  }
  .callout b{color:var(--ink)}
  .run{font-family:var(--font-mono); font-size:.78rem; color:var(--patina); margin:-.4rem 0 1rem}

  footer{padding:40px 0 60px; border-top:1px solid var(--line-soft); color:var(--ink-faint); font-size:.85rem}
  footer .wrap{display:flex; justify-content:space-between; flex-wrap:wrap; gap:1rem}
  footer .mono{font-family:var(--font-mono)}
  ::selection{background:var(--oxide); color:#fff}
  :focus-visible{outline:2px solid var(--oxide); outline-offset:2px; border-radius:3px}
</style>
</head><body><header class="top">
  <div class="wrap">
    <div class="brand"><a href="index.html"><span class="mark"></span>millwright</a><span class="ver">the&nbsp;guide</span></div>
    <nav>
      <a href="index.html">design brief</a>
      <a href="#install">install</a>
      <a href="#pipeline">pipelines</a>
      <a href="#operate">operate</a>
      <a href="#reproducible">reproducible</a>
      <a class="repo" href="https://crates.io/crates/millwright">crates.io ↗</a>
      <a class="repo" href="https://github.com/mi7plus/millwright">GitHub ↗</a>
    </nav>
  </div>
</header>

<main>
  <div class="wrap">
    <div class="hero">
      <div class="eyebrow">The guide</div>
      <h1>Data in, monitored<br>service out.</h1>
      <div class="sub">the hands-on tutorial · one data model, one contract, one pipeline</div>
      <p class="lede">The <a href="index.html">design brief</a> is the <b>why</b>. This is the <b>how</b>: a run through the whole lifecycle — from a raw table of <code class="inl">f64</code> to a served, drift-monitored model — where every capability is a cargo feature over one crate. Every snippet below is real API, mirrored from the runnable programs in <code class="inl">examples/</code>.</p>
    </div>
  </div>

  <!-- ===================== INSTALL ===================== -->
  <section id="install">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 01 · install</div>
        <h2>Pull only what you need.</h2>
        <p class="muted"><code class="inl">default</code> is a lean, useful core — the smartcore backend, preprocessing, model selection, and ensembles. Everything past it is feature-gated: a serving binary never compiles SHAP; a notebook never compiles <code class="inl">axum</code>.</p>
      </div>
<pre><span class="c"># Cargo.toml — the default core</span>
millwright = <span class="s">"0.1"</span>

<span class="c"># just the spine</span>
millwright = { version = <span class="s">"0.1"</span>, default-features = <span class="k">false</span>, features = [<span class="s">"smartcore-backend"</span>] }

<span class="c"># the whole lifecycle</span>
millwright = { version = <span class="s">"0.1"</span>, features = [<span class="s">"full"</span>] }</pre>
      <p class="muted">One import brings the framework into scope, and the lifecycle stages map to features:</p>
      <div class="tablewrap">
        <table>
          <thead><tr><th>Feature</th><th>Adds</th></tr></thead>
          <tbody>
            <tr class="core"><td><span class="feat">smartcore-backend</span><span class="badge">default</span></td><td>RandomForest · LinearRegression</td></tr>
            <tr class="core"><td><span class="feat">preprocessing</span><span class="badge">default</span></td><td>Smote · RandomOverSampler (imputers/scalers/encoders are core)</td></tr>
            <tr class="core"><td><span class="feat">model-selection</span><span class="badge">default</span></td><td>KFold · StratifiedKFold · GridSearch · RandomSearch · metrics</td></tr>
            <tr class="core"><td><span class="feat">ensemble</span><span class="badge">default</span></td><td>Voting · Bagging · Stacking</td></tr>
            <tr><td><span class="feat">eda</span></td><td>Table (polars CSV/Parquet ingest) · Profile (typed EDA)</td></tr>
            <tr><td><span class="feat">linfa-backend</span></td><td>KMeans · GaussianMixture · Dbscan · Pca</td></tr>
            <tr><td><span class="feat">hpo</span></td><td>BayesSearch (TPE) over a SearchSpace</td></tr>
            <tr><td><span class="feat">diagnostics</span></td><td>OLS Diagnostics: VIF · residuals · Cook's distance</td></tr>
            <tr><td><span class="feat">explain</span></td><td>Explainer (SHAP) · permutation_importance</td></tr>
            <tr><td><span class="feat">calibration</span></td><td>PlattScaling · IsotonicRegression · reliability_curve</td></tr>
            <tr><td><span class="feat">anomaly</span></td><td>Mahalanobis · KnnScore outlier detectors</td></tr>
            <tr><td><span class="feat">viz</span></td><td>ROC / residual SVG figures</td></tr>
            <tr><td><span class="feat">onnx</span></td><td>export_onnx · InferenceModel (tract)</td></tr>
            <tr><td><span class="feat">registry</span></td><td>versioned model Registry</td></tr>
            <tr><td><span class="feat">monitor</span></td><td>DriftMonitor (PSI)</td></tr>
            <tr><td><span class="feat">serve</span></td><td>Server — POST /predict, GET /metrics</td></tr>
            <tr><td><span class="feat">timeseries</span></td><td>AutoArima forecaster</td></tr>
            <tr><td><span class="feat">incremental</span></td><td>IncrementalLinear (partial_fit)</td></tr>
            <tr><td><span class="feat">automl</span></td><td>AutoML search</td></tr>
            <tr><td><span class="feat">python</span></td><td>the <code class="inl">pip install millwright</code> package</td></tr>
          </tbody>
        </table>
      </div>

      <div class="eyebrow n" style="margin-top:1.2rem">On this page</div>
      <div class="toc">
        <a href="#ingest"><span class="num"></span><span>Ingest &amp; EDA</span></a>
        <a href="#data"><span class="num">02</span><span>Frame &amp; Dataset</span></a>
        <a href="#contract"><span class="num">03</span><span>The four traits</span></a>
        <a href="#pipeline"><span class="num">04</span><span>Pipelines &amp; params</span></a>
        <a href="#prep"><span class="num">05</span><span>Preprocess &amp; balance</span></a>
        <a href="#select"><span class="num">06</span><span>CV &amp; search</span></a>
        <a href="#ensemble"><span class="num">07</span><span>Ensembles</span></a>
        <a href="#linfa"><span class="num">08</span><span>A second backend</span></a>
        <a href="#insight"><span class="num">09</span><span>Evaluate &amp; explain</span></a>
        <a href="#onnx"><span class="num">10</span><span>ONNX in &amp; out</span></a>
        <a href="#operate"><span class="num">11</span><span>Registry, drift, serve</span></a>
        <a href="#special"><span class="num">12</span><span>Time series &amp; streaming</span></a>
        <a href="#automl"><span class="num">13</span><span>AutoML</span></a>
        <a href="#python"><span class="num">14</span><span>Python</span></a>
        <a href="#reproducible"><span class="num">15</span><span>Reproducibility</span></a>
      </div>
    </div>
  </section>

  <!-- ===================== INGEST ===================== -->
  <section id="ingest">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">The front of the lifecycle</div>
        <h2>Ingest &amp; explore: <span class="mono">Table</span> and <span class="mono">Profile</span>.</h2>
        <p class="muted">Behind the <code class="inl">eda</code> feature, a polars-backed <code class="inl">Table</code> reads real CSV/Parquet — strings, categories, dates, booleans, nulls — and a <code class="inl">Profile</code> reports it and drafts the preprocessing. <code class="inl">Frame</code> stays the numeric boundary; <code class="inl">Table</code> is the typed world <em>in front of</em> it.</p>
      </div>
<pre><span class="k">let</span> table = <span class="f">Table</span>::from_csv(<span class="s">"customers.csv"</span>)?;   <span class="c">// or ::from_parquet(…)</span>

<span class="c">// a typed profile — not just an HTML blob</span>
<span class="k">let</span> profile = <span class="f">Profile</span>::of_with_target(&amp;table, <span class="s">"churned"</span>)?;
<span class="f">println!</span>(<span class="s">"{}"</span>, profile.summary());
<span class="k">for</span> alert <span class="k">in</span> profile.alerts() {
    <span class="f">println!</span>(<span class="s">"{alert}"</span>);   <span class="c">// "[city] categorical (3 levels) → OneHotEncoder"</span>
}
profile.to_html(<span class="s">"eda_report.html"</span>)?;             <span class="c">// a shareable report</span></pre>
      <p class="muted">Because Millwright owns EDA <em>and</em> the pipeline, the profile drafts the starting preprocessing from its own findings — the loop scikit-learn can't close:</p>
<pre><span class="c">// lower the typed table to the numeric world</span>
<span class="k">let</span> train = table.into_dataset(<span class="s">"churned"</span>)?;   <span class="c">// categoricals encoded, nulls → NaN</span>

<span class="c">// EDA drafts the pipeline; you just add the model</span>
<span class="k">let mut</span> pipe = profile.suggest_pipeline()        <span class="c">// impute · encode · scale, from the alerts</span>
    .estimator(<span class="s">"rf"</span>, <span class="f">RandomForest</span>::new());
pipe.fit(&amp;train)?;</pre>
      <p class="run">cargo run --example explore --features "eda smartcore-backend"</p>
    </div>
  </section>

  <!-- ===================== DATA ===================== -->
  <section id="data">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 02 · data</div>
        <h2>The boundary type: <span class="mono">Frame</span> and <span class="mono">Dataset</span>.</h2>
        <p class="muted">Everything the <em>public</em> API speaks is a <code class="inl">Frame</code>: a contiguous, row-major <code class="inl">f64</code> buffer plus a schema. It is what lets a linfa model and a smartcore <code class="inl">DenseMatrix</code> meet in one signature without your code ever naming their array versions — each backend converts <code class="inl">Frame</code> ⇄ its native type inside the adapter only. A <code class="inl">Dataset</code> pairs a frame with a target.</p>
      </div>
<pre><span class="k">use</span> millwright::prelude::*;

<span class="k">let</span> x = <span class="f">Frame</span>::from_rows(
    <span class="f">vec!</span>[<span class="f">vec!</span>[<span class="k">0.0</span>, <span class="k">0.1</span>], <span class="f">vec!</span>[<span class="k">0.4</span>, <span class="k">0.2</span>], <span class="f">vec!</span>[<span class="k">9.0</span>, <span class="k">9.1</span>], <span class="f">vec!</span>[<span class="k">9.4</span>, <span class="k">8.7</span>]],
    <span class="f">vec!</span>[<span class="s">"a"</span>.into(), <span class="s">"b"</span>.into()],
)?;
<span class="k">assert_eq!</span>(x.shape(), (<span class="k">4</span>, <span class="k">2</span>));        <span class="c">// (rows, cols)</span>

<span class="k">let</span> train = <span class="f">Dataset</span>::new(x.clone(), <span class="f">vec!</span>[<span class="k">0.0</span>, <span class="k">0.0</span>, <span class="k">1.0</span>, <span class="k">1.0</span>])?;
<span class="k">let</span> _features = train.features();  <span class="c">// &Frame</span>
<span class="k">let</span> _target   = train.target();    <span class="c">// &[f64]</span></pre>
      <p class="tiny">The task — classification vs. regression — is inferred from the target: an all-integral target is class labels, anything else is regression.</p>
    </div>
  </section>

  <!-- ===================== CONTRACT ===================== -->
  <section id="contract">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 03 · the contract</div>
        <h2>Four object-safe traits.</h2>
        <p class="muted">The whole core is four traits. Object-safe means a <code class="inl">Pipeline</code> can hold a heterogeneous <code class="inl">Vec&lt;Box&lt;dyn …&gt;&gt;</code> — everything composes because everything speaks the same contract.</p>
      </div>
      <div class="tablewrap">
        <table>
          <thead><tr><th>Trait</th><th>Method</th><th>Meaning</th></tr></thead>
          <tbody>
            <tr><td><span class="feat">Transformer</span></td><td>fit(&amp;mut, &amp;Frame) → transform(&amp;Frame)</td><td>learn column stats, reshape features</td></tr>
            <tr><td><span class="feat">Estimator</span></td><td>fit(&amp;mut, &amp;Dataset)</td><td>learn a model from features + target</td></tr>
            <tr><td><span class="feat">Predictor</span></td><td>predict(&amp;Frame) → Vec&lt;f64&gt;</td><td>one prediction per row</td></tr>
            <tr><td><span class="feat">ProbaPredictor</span></td><td>predict_proba(&amp;Frame) → Frame</td><td>class probabilities</td></tr>
          </tbody>
        </table>
      </div>
      <p class="muted">A blanket <code class="inl">Model</code> ties <code class="inl">Estimator + Predictor</code> together; a blanket <code class="inl">Evaluate</code> gives every predictor a <code class="inl">.evaluate(&amp;test)</code>. You rarely name these — you compose the types that implement them.</p>
    </div>
  </section>

  <!-- ===================== PIPELINE ===================== -->
  <section id="pipeline">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 04 · compose</div>
        <h2>Pipelines: compose, then tune by path.</h2>
        <p class="muted">A <code class="inl">Pipeline</code> is named transformer steps plus one final estimator, as a single object that is itself a <code class="inl">Model</code>. Steps are addressable by name, so you tune a parameter deep in the chain by path — the scikit-learn <code class="inl">"step__param"</code> convention.</p>
      </div>
<pre><span class="k">let mut</span> pipe = <span class="f">Pipeline</span>::new()
    .step(<span class="s">"scale"</span>, <span class="f">StandardScaler</span>::new())
    .estimator(<span class="s">"rf"</span>, <span class="f">RandomForest</span>::new());

pipe.set_param(<span class="s">"rf__n_trees"</span>, <span class="f">ParamValue</span>::Int(<span class="k">50</span>))?;   <span class="c">// tune by path</span>
pipe.set_param(<span class="s">"rf__max_depth"</span>, <span class="f">ParamValue</span>::Int(<span class="k">4</span>))?;

pipe.fit(&amp;train)?;                <span class="c">// fit transforms, then the estimator</span>
<span class="k">let</span> preds = pipe.predict(&amp;x)?;   <span class="c">// replays fitted transforms, then the model</span></pre>
      <p class="run">cargo run --example spine</p>
      <p class="tiny">That path addressing is exactly what lets a <em>search</em> reach any parameter anywhere in the chain. Pipelines nest, so a pipeline can be a step in another.</p>
    </div>
  </section>

  <!-- ===================== PREP ===================== -->
  <section id="prep">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 05 · prepare</div>
        <h2>Preprocess &amp; balance.</h2>
        <p class="muted">The core transformers need no extra dependencies. <em>Balancers</em> (feature <code class="inl">preprocessing</code>) are train-time only — they resample during <code class="inl">fit</code> and are skipped at predict time, so they never distort inference.</p>
      </div>
<pre><span class="k">let</span> pipe = <span class="f">Pipeline</span>::new()
    .step(<span class="s">"impute"</span>, <span class="f">SimpleImputer</span>::median())   <span class="c">// or ::mean() / ::constant(0.0)</span>
    .step(<span class="s">"scale"</span>,  <span class="f">StandardScaler</span>::new())      <span class="c">// or MinMaxScaler::new()</span>
    .step(<span class="s">"encode"</span>, <span class="f">OneHotEncoder</span>::infer())      <span class="c">// or ::columns(["city"])</span>
    .balance(<span class="f">Smote</span>::new().k_neighbors(<span class="k">3</span>).random_state(<span class="k">0</span>))  <span class="c">// train-time only</span>
    .estimator(<span class="s">"rf"</span>, <span class="f">RandomForest</span>::new());</pre>
    </div>
  </section>

  <!-- ===================== SELECT ===================== -->
  <section id="select">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 06 · select</div>
        <h2>Cross-validate &amp; search — over the whole pipeline.</h2>
        <p class="muted">Search runs over the entire pipeline, cross-validated, tuning by path. <code class="inl">RandomSearch</code> swaps the grid for random draws; with <code class="inl">hpo</code>, <code class="inl">BayesSearch</code> runs TPE and returns the <em>same</em> <code class="inl">SearchResult</code> — one API, three strategies.</p>
      </div>
<pre><span class="k">use</span> millwright::grid;

<span class="k">let</span> search = <span class="f">GridSearch</span>::new(pipe, <span class="f">grid!</span> { <span class="s">"rf__max_depth"</span> =&gt; [<span class="k">2</span>, <span class="k">4</span>, <span class="k">8</span>] })
    .cv(<span class="f">StratifiedKFold</span>::new(<span class="k">4</span>))
    .scoring(<span class="f">Metric</span>::F1)
    .fit(&amp;train)?;

<span class="k">for</span> (params, score) <span class="k">in</span> search.leaderboard() {
    <span class="f">println!</span>(<span class="s">"  {score:.3}  {params:?}"</span>);
}
<span class="f">println!</span>(<span class="s">"best F1 = {:.3}"</span>, search.best_score());
<span class="k">let</span> preds = search.predict(&amp;probe)?;   <span class="c">// the refit best model</span></pre>
      <p class="run">cargo run --example workflow · cargo run --example backends --features "linfa-backend hpo"</p>
    </div>
  </section>

  <!-- ===================== ENSEMBLE ===================== -->
  <section id="ensemble">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 07 · combine</div>
        <h2>Ensembles — even across backends.</h2>
        <p class="muted">Because every model is a <code class="inl">Predictor</code>, combining models is just another <code class="inl">Predictor</code> that holds several — no new machinery, and it works across backends. An ensemble <em>is</em> an <code class="inl">Estimator</code>, so you can search a member straight through it.</p>
      </div>
<pre><span class="c">// soft (mean-probability) vote across two forests</span>
<span class="k">let mut</span> vote = <span class="f">Voting</span>::soft()
    .add(<span class="s">"rf_shallow"</span>, <span class="f">RandomForest</span>::new().max_depth(<span class="k">2</span>))
    .add(<span class="s">"rf_deep"</span>,    <span class="f">RandomForest</span>::new().max_depth(<span class="k">8</span>));
vote.fit(&amp;train)?;

<span class="c">// stacking: a meta-learner on leak-free out-of-fold base predictions</span>
<span class="k">let mut</span> stack = <span class="f">Stacking</span>::meta(<span class="f">RandomForest</span>::new().n_trees(<span class="k">50</span>))
    .base(<span class="s">"rf"</span>,  <span class="f">RandomForest</span>::new().n_trees(<span class="k">30</span>))
    .base(<span class="s">"rf2"</span>, <span class="f">RandomForest</span>::new().max_depth(<span class="k">3</span>))
    .cv(<span class="f">StratifiedKFold</span>::new(<span class="k">4</span>));   <span class="c">// folds from the CV engine → leak-free</span>
stack.fit(&amp;train)?;</pre>
    </div>
  </section>

  <!-- ===================== LINFA ===================== -->
  <section id="linfa">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 08 · a second backend</div>
        <h2>linfa — through the same <span class="mono">Frame</span>.</h2>
        <p class="muted">The <code class="inl">linfa-backend</code> feature adds unsupervised models through the same boundary type — the proof that the two-<code class="inl">ndarray</code>-worlds problem is settled by design. Clusterers implement a <code class="inl">Clusterer</code> contract; <code class="inl">Pca</code> is a <code class="inl">Transformer</code>.</p>
      </div>
<pre><span class="k">let mut</span> km = <span class="f">KMeans</span>::new(<span class="k">2</span>);
km.fit(&amp;x)?;
<span class="f">println!</span>(<span class="s">"k-means labels: {:?}"</span>, km.predict(&amp;x)?);

<span class="k">let</span> dbscan = <span class="f">Dbscan</span>::new(<span class="k">3</span>).tolerance(<span class="k">1.0</span>);
<span class="f">println!</span>(<span class="s">"dbscan: {:?}"</span>, dbscan.fit_predict(&amp;x)?);

<span class="k">let mut</span> pca = <span class="f">Pca</span>::new(<span class="k">1</span>);
<span class="k">let</span> reduced = pca.fit_transform(&amp;x)?;   <span class="c">// a Frame with fewer columns</span></pre>
    </div>
  </section>

  <!-- ===================== INSIGHT ===================== -->
  <section id="insight">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 09 · insight</div>
        <h2>Evaluate, diagnose, explain, visualize.</h2>
        <p class="muted">Any predictor scores itself on a labelled set (core). <code class="inl">explain</code> adds SHAP and permutation importance; <code class="inl">diagnostics</code> adds OLS VIF/residuals/influence; <code class="inl">viz</code> renders self-contained SVGs.</p>
      </div>
<pre><span class="k">let mut</span> rf = <span class="f">RandomForest</span>::new().n_trees(<span class="k">60</span>);
rf.fit(&amp;train)?;
<span class="f">print!</span>(<span class="s">"{}"</span>, rf.evaluate(&amp;test)?);   <span class="c">// accuracy / precision / recall / F1</span>

<span class="c">// explain (feature = "explain")</span>
<span class="k">let</span> shap = rf.explain(&amp;<span class="f">Explainer</span>::kernel().nsamples(<span class="k">80</span>), test.features())?;
<span class="k">let</span> perm = <span class="f">permutation_importance</span>(&amp;rf, &amp;test, <span class="k">8</span>, <span class="k">0</span>)?;

<span class="c">// diagnostics (feature = "diagnostics") · viz (feature = "viz")</span>
<span class="k">let</span> diag = <span class="f">Diagnostics</span>::of(&amp;reg)?;
<span class="f">println!</span>(<span class="s">"R² = {:.4}, VIF = {:?}"</span>, diag.r_squared(), diag.vif());
<span class="k">let</span> auc = viz::roc_svg(test.target(), &amp;scores, <span class="s">"roc.svg"</span>, (<span class="k">520</span>, <span class="k">420</span>))?;</pre>
      <p class="run">cargo run --example insight --features "diagnostics explain viz"</p>
    </div>
  </section>

  <!-- ===================== ONNX ===================== -->
  <section id="onnx">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 10 · portability</div>
        <h2>ONNX in and out.</h2>
        <p class="muted">With <code class="inl">onnx</code>, any model — or a whole pipeline — exports to one <code class="inl">.onnx</code> file. Whole-pipeline export folds leading affine scalers into the estimator's graph: raw features in, predictions out. <code class="inl">InferenceModel::load</code> runs any ONNX file back through tract.</p>
      </div>
<pre><span class="k">let mut</span> pipe = <span class="f">Pipeline</span>::new()
    .step(<span class="s">"scale"</span>, <span class="f">StandardScaler</span>::new())
    .estimator(<span class="s">"lr"</span>, <span class="f">LinearRegression</span>::new());
pipe.fit(&amp;train)?;
<span class="k">let</span> native = pipe.predict(&amp;probe)?;

pipe.export_onnx(<span class="s">"pipeline.onnx"</span>)?;                 <span class="c">// scaler + model, one graph</span>
<span class="k">let</span> model = <span class="f">InferenceModel</span>::load(<span class="s">"pipeline.onnx"</span>)?;
<span class="k">let</span> via_onnx = model.predict(&amp;probe)?;              <span class="c">// matches `native`</span></pre>
      <p class="tiny">Linear/affine/pipeline graphs run inside tract for a full round-trip. A random forest exports to a valid ONNX-ML tree-ensemble artifact for external runtimes (onnxruntime); tract implements NN ops, not the ONNX-ML tree ops.</p>
    </div>
  </section>

  <!-- ===================== OPERATE ===================== -->
  <section id="operate">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 11 · operate</div>
        <h2>Registry, drift, serving.</h2>
        <p class="muted">Where Millwright runs past where scikit-learn stops. The <code class="inl">registry</code> versions the ONNX artifact (content-addressed, with a reference distribution and movable tags); <code class="inl">monitor</code> watches the prediction stream for PSI drift; <code class="inl">serve</code> exposes a validated endpoint that feeds the monitor.</p>
      </div>
<pre><span class="k">let</span> reg = <span class="f">Registry</span>::local(<span class="s">"./models"</span>);
<span class="k">let</span> v1 = reg.register(<span class="s">"demand"</span>, &amp;model, <span class="f">Metadata</span> {
    metrics: <span class="f">vec!</span>[(<span class="s">"r2"</span>.into(), <span class="k">1.0</span>)],
    reference: reference.clone(),   <span class="c">// the distribution drift watches against</span>
    note: <span class="s">"baseline"</span>.into(),
})?;
reg.tag(<span class="s">"demand"</span>, &amp;v1.id, <span class="s">"prod"</span>)?;
<span class="k">let</span> reverted = reg.rollback(<span class="s">"demand"</span>, <span class="s">"prod"</span>)?;   <span class="c">// revert in one line</span>

<span class="c">// serve the prod artifact, watching for drift on every request</span>
<span class="f">Server</span>::from_onnx(reg.onnx_path(<span class="s">"demand"</span>, <span class="s">"prod"</span>)?)?
    .route(<span class="s">"/predict"</span>)
    .with_monitor(<span class="f">DriftMonitor</span>::psi(&amp;reference)?)
    .serve(<span class="s">"0.0.0.0:8080"</span>).<span class="k">await</span>?;         <span class="c">// POST /predict, GET /metrics</span></pre>
      <p class="run">cargo run --example operations --features "onnx registry monitor serve"</p>
    </div>
  </section>

  <!-- ===================== SPECIALIZED ===================== -->
  <section id="special">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 12 · specialized shapes</div>
        <h2>Time series &amp; out-of-core.</h2>
        <p class="muted">Same contract, different data shapes — each gets its own trait. These two crates pin <code class="inl">ndarray 0.15</code> while the rest of the stack uses <code class="inl">0.16</code>; Cargo links both and converts only inside the adapters — the "two <code class="inl">ndarray</code> worlds," exercised for real.</p>
      </div>
<pre><span class="c">// time series (feature = "timeseries")</span>
<span class="k">let mut</span> arima = <span class="f">AutoArima</span>::new().max_p(<span class="k">3</span>).max_q(<span class="k">3</span>);
arima.fit(&amp;series)?;                  <span class="c">// &[f64]</span>
<span class="k">let</span> forecast = arima.forecast(<span class="k">6</span>)?;   <span class="c">// six steps ahead</span>

<span class="c">// out-of-core (feature = "incremental") — never holds the whole set in memory</span>
<span class="k">let mut</span> model = <span class="f">IncrementalLinear</span>::with_rate(<span class="k">0.05</span>, <span class="k">0.0</span>);
<span class="k">for</span> batch <span class="k">in</span> batches {
    model.partial_fit(&amp;batch)?;      <span class="c">// one batch at a time</span>
}</pre>
      <p class="run">cargo run --example specialized --features "timeseries incremental"</p>
    </div>
  </section>

  <!-- ===================== AUTOML ===================== -->
  <section id="automl">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 13 · synthesis</div>
        <h2>AutoML — the framework, pointed at itself.</h2>
        <p class="muted">Profiling, preprocessing, CV, search, and ensembling are exactly what an AutoML engine needs — so <code class="inl">AutoML</code> is not a bolt-on, it is the framework orchestrating its own parts. Point it at data and a budget; get a leaderboard and the best <em>deployable</em> model.</p>
      </div>
<pre><span class="k">let</span> result = <span class="f">AutoML</span>::classifier()      <span class="c">// or ::regressor()</span>
    .budget(<span class="f">Budget</span>::trials(<span class="k">20</span>))         <span class="c">// or Budget::minutes(10)</span>
    .metric(<span class="f">Metric</span>::F1)
    .cv(<span class="f">StratifiedKFold</span>::new(<span class="k">5</span>))
    .seed(<span class="k">0</span>)
    .fit(&amp;train)?;

<span class="f">println!</span>(<span class="s">"{}"</span>, result.leaderboard());
<span class="f">println!</span>(<span class="s">"winner: {} (F1 = {:.3})"</span>, result.best_label(), result.best_score());
result.export_onnx(<span class="s">"model.onnx"</span>)?;   <span class="c">// deployable — unlike a TPOT object</span></pre>
      <p class="run">cargo run --example automl --features "automl onnx"</p>
    </div>
  </section>

  <!-- ===================== PYTHON ===================== -->
  <section id="python">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 14 · interop</div>
        <h2>Python — the same engine, a Pythonic API.</h2>
        <p class="muted"><code class="inl">pip install millwright</code> — a Pythonic API over the same Rust engine, shipped as an abi3 wheel built with maturin.</p>
      </div>
<pre><span class="k">import</span> millwright <span class="k">as</span> mw

pipe = mw.<span class="f">Pipeline</span>()
pipe.standard_scaler()
pipe.random_forest(n_trees=<span class="k">100</span>, max_depth=<span class="k">8</span>)

pipe.fit(rows, labels)          <span class="c"># list[list[float]], list[float]</span>
preds = pipe.predict(rows)      <span class="c"># runs the Rust engine</span></pre>
      <p class="run">pip install millwright</p>
      <p class="tiny"><code class="inl">python</code> is deliberately not part of <code class="inl">full</code>: pyo3's <code class="inl">extension-module</code> defers libpython symbols, so a plain <code class="inl">cargo test</code> can't link it. It is built and tested the way it ships — as a wheel.</p>
    </div>
  </section>

  <!-- ===================== REPRODUCIBLE ===================== -->
  <section id="reproducible">
    <div class="wrap">
      <div class="head col">
        <div class="eyebrow">Step 15 · bet on it</div>
        <h2>Reproducibility: pins, lockfile, golden tests, CI.</h2>
        <p class="muted">Millwright assembles young, single-author engine crates. That is its real risk, and Phase 8 owns it directly.</p>
      </div>
      <ul class="clean">
        <li><b>Exact-version pins.</b> Every engine — the ecosystem crates plus the smartcore and linfa families — is pinned to an exact <code class="inl">=x.y.z</code>. A stray <code class="inl">cargo update</code> can never silently move a fragile engine under the stable trait contract. General infrastructure (serde, tokio, axum) stays on caret ranges to avoid forcing conflicts downstream.</li>
        <li><b>Committed <code class="inl">Cargo.lock</code>.</b> The whole ~300-package graph is reproducible; CI builds with <code class="inl">--locked</code>, so a drifted lockfile is a hard error.</li>
        <li><b>Golden-output tests.</b> <code class="inl">tests/golden.rs</code> locks the numeric behaviour of the engines on fixed inputs — exact for the deterministic paths, well-separated class labels for the stochastic ones. An engine bump that moves a number shows up as a diff.</li>
        <li><b>Feature-matrix CI.</b> <code class="inl">fmt</code>, <code class="inl">clippy -D warnings</code>, docs, and the tests across the whole matrix — from <code class="inl">--no-default-features</code> through each feature to <code class="inl">full</code> — plus Windows/macOS, the runnable examples, a benchmark compile-check, a <code class="inl">cargo publish --dry-run</code>, and a maturin wheel. The MSRV (<code class="inl">rust-version</code>) is enforced by cargo for consumers.</li>
      </ul>
<pre><span class="c"># run the whole suite yourself</span>
cargo test --features full
cargo test --locked --no-default-features --features smartcore-backend</pre>
    </div>
  </section>

</main>

<footer>
  <div class="wrap">
    <span class="mono">⚙ millwright — the guide</span>
    <span class="mono"><a href="index.html">design brief</a> · <a href="https://crates.io/crates/millwright">crates.io</a> · <a href="https://docs.rs/millwright">docs.rs</a> · <a href="https://github.com/mi7plus/millwright">GitHub</a></span>
  </div>
</footer>
</body></html>