Engineering lab / Interactive technical portfolio

Synthetic demonstration

Manufacturing ML workbench

Predict batch risk. Inspect mistakes. Decide where to investigate.

Factory and quality workspaces

All records below are fictional; no employer or customer engagement is implied.

Inspect batchPrediction model

Batch nonconformance risk

95.3%

At least 1 of 24 pieces outside diameter limits

Inspection recommendation

Escalate

Add dimensional inspection; retain all quality gates

Inspection threshold / 20.0%

Flag when predicted batch risk ≥ threshold. A score below the threshold never bypasses inspection.

Upcoming fictional batch — no measured outcome. Prediction is a demonstration, not release evidence.

SYN-VS25 / M1 / cutoff before machining

Pre-start inputs

Feature	Recorded input	Timing
Prior tool cuts	168 cuts	Cumulative cuts already completed before this batch
Commanded feed	0.2 mm/rev	Planned machine command
Commanded cutting speed	180 m/min	Planned machine command
Pre-start coolant flow	6.2 L/min	Reading taken before machining
Pre-start coolant temperature	26.4 °C	Reading taken before machining
Commanded setup offset	8 µm	Commanded preset; not a first-off diameter measurement
Incoming material hardness	218 HB	Incoming measurement recorded before machining

Edit a hypothetical batch

Editing forks an unobserved scenario. It does not change the training history or attach historical outcomes to altered inputs.

Prior tool cuts / cuts

Commanded feed / mm/rev

Commanded cutting speed / m/min

Pre-start coolant flow / L/minPre-start coolant flow missing

Pre-start coolant temperature / °CPre-start coolant temperature missing

Commanded setup offset / µm

Incoming material hardness / HBIncoming material hardness missing

Association / not a root-cause finding

How the model arrived here

Commanded feed / mm/rev: +2.246 log-odds relative to the training mean.
Pre-start coolant flow / L/min: +1.943 log-odds relative to the training mean.
Commanded cutting speed / m/min: +1.921 log-odds relative to the training mean.
Prior tool cuts / cuts: +1.674 log-odds relative to the training mean.

These are fitted model contributions or tree rules. They do not establish causation. Missing readings use training medians plus indicators; no final diameter or downstream disposition is an input.

Prediction → inspection support → measured evidence

The target is a batch event, not a unit defect percentage. Containment and release still require physical evidence and authorized decisions.

Final holdout / 144 complete batches / 23 positive batches

Test what generalizes

Current threshold and costs explore this fixed holdout. The validation-selected reference stays frozen: Logistic regression at 20.0%. Adjustments after inspecting holdout outcomes are not a new unseen benchmark.

Recall / positives detected

39.1%

9 / 23 positive batches

Precision / flagged positives

90.0%

9 / 10 flagged batches; undefined if none

Additional inspection

10 / 144

Flagged batches / all holdout batches

Assumed cost proxy

$2,700

Flagged × inspection + missed positives × exposure

True positives

False positives

False negatives

True negatives

120

Model	Precision	Recall	TP / FP / FN / TN	Brier ↓	Log loss ↓	Cost proxy
Prevalence baseline	Undefined	0.0%	0 / 0 / 23 / 121	0.1441	0.5010	$4,140
Logistic regression	90.0%	39.1%	9 / 1 / 14 / 120	0.1054	0.4076	$2,700
Decision tree	64.3%	39.1%	9 / 5 / 14 / 116	0.1178	0.8379	$2,772

Brier measures squared probability error. Log loss strongly penalizes confident mistakes. Smaller is better; neither alone proves reliable calibration. Models were chosen on validation Brier, not holdout scores.

Frozen reference results and model selection

Model	Validation Brier	Holdout Brier	Holdout recall	TP / FP / FN / TN
Prevalence baseline	0.0342	0.1441	0.0%	0 / 0 / 23 / 121
Logistic regression	0.0148	0.1054	39.1%	9 / 1 / 14 / 120
Decision tree	0.0268	0.1178	39.1%	9 / 5 / 14 / 116

All frozen confusion counts use the same validation-selected 20% threshold and $18 / $180 assumptions. Only 5 validation batches were positive, so model and threshold selection is uncertain. No refit on validation or holdout.

Candidate	Parameters	Validation Brier
logistic	C=0.1	0.01838
logistic	C=1	0.01477
logistic	C=10	0.01504
tree	depth=3; minimum leaf=16	0.02947
tree	depth=3; minimum leaf=32	0.02684
tree	depth=4; minimum leaf=16	0.02698
tree	depth=4; minimum leaf=32	0.02684

Calibration / five fixed bins

Are predicted probabilities credible?

Probability bin	Batches	Mean prediction	Observed positives
0.0%–20.0% upper exclusive	134	1.4%	10.4%
20.0%–40.0% upper exclusive	5	32.6%	80.0%
40.0%–60.0% upper exclusive	2	54.1%	100.0%
60.0%–80.0% upper exclusive	1	78.7%	100.0%
80.0%–100.0% inclusive	2	84.9%	100.0%

A well-calibrated model lies near the diagonal. Bin size is shown; tiny bins are unstable. These are raw probabilities, with no calibration fit on the holdout.

Inspection effort / missed positive batches

What does a threshold trade?

Additional inspection cost / $ per flagged batchMissed positive-batch exposure / $

Assumptions, not financial results: additional inspection detects every flagged defect; no credit for detection by standard inspection. False negatives are missed flags, not demonstrated customer escapes. Fixed overhead and actual disposition costs are excluded.

Threshold counts and assumed costs / 49 points

Threshold	Precision	Recall	Flagged	Missed positive	Cost proxy
2.0%	56.3%	78.3%	32 / 144	5 / 23	$1,476
4.0%	63.0%	73.9%	27 / 144	6 / 23	$1,566
6.0%	70.0%	60.9%	20 / 144	9 / 23	$1,980
8.0%	78.6%	47.8%	14 / 144	12 / 23	$2,412
10.0%	84.6%	47.8%	13 / 144	12 / 23	$2,394
12.0%	84.6%	47.8%	13 / 144	12 / 23	$2,394
14.0%	84.6%	47.8%	13 / 144	12 / 23	$2,394
16.0%	90.9%	43.5%	11 / 144	13 / 23	$2,538
18.0%	90.0%	39.1%	10 / 144	14 / 23	$2,700
20.0%	90.0%	39.1%	10 / 144	14 / 23	$2,700
22.0%	90.0%	39.1%	10 / 144	14 / 23	$2,700
24.0%	88.9%	34.8%	9 / 144	15 / 23	$2,862
26.0%	88.9%	34.8%	9 / 144	15 / 23	$2,862
28.0%	88.9%	34.8%	9 / 144	15 / 23	$2,862
30.0%	87.5%	30.4%	8 / 144	16 / 23	$3,024
32.0%	87.5%	30.4%	8 / 144	16 / 23	$3,024
34.0%	87.5%	30.4%	8 / 144	16 / 23	$3,024
36.0%	87.5%	30.4%	8 / 144	16 / 23	$3,024
38.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
40.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
42.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
44.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
46.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
48.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
50.0%	100.0%	21.7%	5 / 144	18 / 23	$3,330
52.0%	100.0%	17.4%	4 / 144	19 / 23	$3,492
54.0%	100.0%	17.4%	4 / 144	19 / 23	$3,492
56.0%	100.0%	17.4%	4 / 144	19 / 23	$3,492
58.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
60.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
62.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
64.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
66.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
68.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
70.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
72.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
74.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
76.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
78.0%	100.0%	13.0%	3 / 144	20 / 23	$3,654
80.0%	100.0%	8.7%	2 / 144	21 / 23	$3,816
82.0%	100.0%	8.7%	2 / 144	21 / 23	$3,816
84.0%	100.0%	4.3%	1 / 144	22 / 23	$3,978
86.0%	100.0%	4.3%	1 / 144	22 / 23	$3,978
88.0%	Undefined	0.0%	0 / 144	23 / 23	$4,140
90.0%	Undefined	0.0%	0 / 144	23 / 23	$4,140
92.0%	Undefined	0.0%	0 / 144	23 / 23	$4,140
94.0%	Undefined	0.0%	0 / 144	23 / 23	$4,140
96.0%	Undefined	0.0%	0 / 144	23 / 23	$4,140
98.0%	Undefined	0.0%	0 / 144	23 / 23	$4,140

Chronology / hidden from model inputs

Investigate the changed condition

Blue line: predicted batch risk. Dashed line: inspection threshold. × marks a batch with a final dimensional nonconformance.

At batch 625, the generator introduces an unobserved +18 µm thermal offset and a flow reading that overstates actual coolant by 1.2 L/min. Missingness also increases. The model sees the pre-start readings, never the hidden condition. The highlighted segment illustrates failure under a changed process.

Segment	Batches	Positive batches	Recall	False negatives	Brier
nominal / 577–624	48	3	100.0%	0	0.0135
changed / 625–720	96	20	30.0%	14	0.1514

Outcome filter

14 / 144 heldout batches

Risk	Nonconforming / 24 pieces	Outcome	Condition
17.2%	1 / 24	FN	changed
6.5%	12 / 24	FN	changed
1.2%	3 / 24	FN	changed
4.8%	1 / 24	FN	changed
0.1%	1 / 24	FN	changed
3.1%	2 / 24	FN	changed
6.6%	1 / 24	FN	changed
5.8%	3 / 24	FN	changed
1.8%	1 / 24	FN	changed
4.2%	2 / 24	FN	changed
0.4%	1 / 24	FN	changed
0.7%	2 / 24	FN	changed
15.6%	4 / 24	FN	changed
7.3%	3 / 24	FN	changed

Independent synthetic demonstration / actual trained models

Engineering case record & reproducibility

Problem & scope: Prioritize dimensional inspection for upcoming valve batches in this fictional cooling-skid operation. Independent 720-batch learning history; 17,280 synthetic final diameters.
Responsibility & approach: This demonstration presents data generation, feature/cutoff design, offline model training, evaluation and browser inference. It is separate from documented professional work and AutoOps AI.
Decisions & implementation: Use whole chronological batches, training-only median imputation and missing indicators. Scale logistic inputs on training data only. Compare a prevalence baseline, L2 logistic regression and a shallow tree to handle nonlinear relationships. Export fitted parameters for local inference; no remote AI API.
Measured synthetic result: Frozen logistic policy: 9 true positives, 1 false positive, 14 false negatives and 120 true negatives across 144 heldout batches. Precision 90%; recall 39.1%. These are synthetic evaluation outcomes, not real customer results.
Lessons & next evidence: Few positive validation batches and a changed process limit confidence. Model associations do not diagnose a cause. Real use needs representative, permissioned data, validated measurement systems, prospective evaluation, drift monitoring and accountable inspection decisions.

Split	Batch sequence	Batches / pieces	Positive batches	Missing input cells
train	1–432	432 / 10368	26	69
validation	433–576	144 / 3456	5	14
holdout	577–720	144 / 3456	23	18

All 24 pieces remain in their parent batch. Final diameters, nonconforming counts, target, batch sequence, split, hidden condition, NCR and FAT dispositions are excluded from prediction inputs. The 32-run improvement experiment and original SPC records remain separate.

Download ML method & model card Download dataset manifest Download 720 batches / CSV Download 17,280 pieces / CSV Download trained model parameters Download frozen evaluation Download reference predictions Download training pipeline Download pinned training requirements

Method references: training-only preprocessing and probability calibration from scikit-learn documentation.

Compare with the controlled experiment Investigate process signals Explore data & product development

Factory planning scenario and reset

SYN-OPS-01 / Factory overview

factory-1.0 · Synthetic operation

Planning scenarioRequested skids / week

Plan: 8 skids / 54.8 h assumed work. Baseline: 8 skids / 54.8 h.

This tab retains your scenario and demo reviews. Reset factory restores the original state.

One connected perspective

Physical requirements become data. Data becomes evidence. Evidence informs an operational decision. These demonstrations connect my manufacturing and quality foundation with the systems thinking used in enterprise transformation and product engineering.

Professional context

Explore my documented manufacturing experience, SAP readiness work and data and product development. The synthetic lab is separate from those professional records.

Download technical README