INDEPENDENT SYNTHETIC CASE STUDY · FRANCE · 2026

Specialty Care AI Control Tower

An end-to-end responsible AI product combining governed data engineering, exploratory analysis, supervised and unsupervised learning, deep-learning forecasting, Bayesian inference, controlled retrieval, constrained optimization and production-ready software.

800 HCPs48 territories104 weeksPython + SparkLightGBM + PyTorchFastAPI + DockerPower BI + Tableau

Explore the results Read technical report

All records, outcomes and notes are synthetic. No patient-identifying data are used.

System snapshot

v1.0.0

15.0%

Forecast WAPE

0.801

Propensity AUC

94.9%

NLP macro-F1

42.4

Expected starts

National weekly starts

Actual vs LightGBM champion · 12-week holdout

15.0% WAPE

Actual LightGBM champion80% interval

Holdout weeks 92–10386.6% empirical interval coverage

Data contracts

PASS

Model monitoring

PASS

Human review

HUMAN REVIEW REQUIRED

EXECUTIVE CONTROL TOWER

One governed analytical chain, from source data to human-reviewed decisions

The system is designed as a product rather than a notebook: source contracts feed bronze, silver and gold layers; versioned models generate forecasts, propensity scores, segments and note routes; a mixed-integer optimizer produces a capacity-feasible plan; every result is monitored and traceable.

83,200

HCP-week rows

gold table

100.000%

Gold completeness

160 duplicates removed

Actionable segments

silhouette 0.279

0.801

Propensity ROC AUC

2.20x top-decile lift

15.0%

LightGBM Poisson WAPE

R² 0.771

86.6%

Nominal 80% coverage

adaptive q = 0.309

95.0%

Retrieval recall@1

100% JSON conformance

€30,000

Optimized budget

42.4 expected starts

The project demonstrates the complete AI lifecycle: business translation, information-system understanding, data quality, EDA, supervised and unsupervised learning, deep learning, Bayesian statistics, optimization, LLM/RAG guardrails, cloud-scale patterns, dashboards, deployment, monitoring and responsible AI.

ANALYTICAL CHAIN

Eight linked stages, one accountable pipeline

Source contracts

CRM, ERP/supply proxies, consent, master data and approved content.

Medallion platform

Bronze replayability, silver quality contracts and gold feature marts.

Exploratory analysis

Quality profiling, seasonality, access, engagement and leakage checks.

Predictive models

LightGBM propensity, LightGBM Poisson and global LSTM.

Uncertainty

Bayesian posteriors and split-conformal forecast intervals.

Controlled AI

TF-IDF classification, approved retrieval, JSON contract and safety escalation.

Optimization

Binary MILP under budget, capacity, contact and consent constraints.

Operations

API, Docker, CI, model registry, monitoring, Excel and BI dashboards.

DATA PLATFORM

Synthetic scale, explicit contracts and an auditable medallion architecture

Architecture

CRM / ERP / consent

Bronze object storage

Silver Spark quality

Gold feature marts

BI + API

ML registry + tests

Forecast / propensity / segments / NLP

MILP + human review

Scope

800

synthetic HCPs

metropolitan regions

synthetic territories

104

weekly periods

74,326

CRM interactions

5,000

non-identifying field notes

Data-quality issues

Completeness: raw vs gold

Source	Grain / count	Role
HCP master	800	profile, territory, potential, consent
HCP-week activity	83,200	engagement, stock, access, starts
CRM interactions	74,326	channel, duration, estimated cost
Field notes	5,000	controlled routing categories
Territory master	48	population proxy and access indices
Approved KB	20	retrieval corpus

Note. Population and physician-demography sources shape broad regional plausibility only. Individual records and outcome relationships are fully synthetic.

UNSUPERVISED LEARNING

Three interpretable profiles selected quantitatively

Seven standardized profile features are clustered with K-means. Candidate k values from 3 to 6 are compared by silhouette score; k=3 is selected. PCA is used only for visualization.

PCA projection · colored by segment

Silhouette by k

Selected: k = 3 (score 0.279).

High-potential omnichannel

Annual starts: 156.4
Contacts / week: 1.24
Portal sessions / week: 1.45
Potential: 0.748
Digital affinity: 0.703
Response posterior: 41.0% [39.9%, 42.1%]

Coordinated, consent-aware omnichannel planning.

Access-constrained / maintain

Annual starts: 103.5
Contacts / week: 0.89
Portal sessions / week: 1.10
Potential: 0.496
Digital affinity: 0.520
Access barrier: 0.397
Response posterior: 37.8% [36.9%, 38.7%]

Local access review matters more than contact volume alone.

Low-engagement / nurture

Annual starts: 74.3
Contacts / week: 0.57
Portal sessions / week: 0.74
Potential: 0.269
Digital affinity: 0.322
Response posterior: 34.7% [33.6%, 35.8%]

Selective, low-cost approved education and follow-up.

Segment labels are assigned after clustering from aggregate centroids. They do not imply clinical value or professional quality.

SUPERVISED LEARNING

Out-of-time propensity scoring with calibration and subgroup review

ROC curve

AUC 0.801

0.801

ROC AUC

0.712

PR AUC

0.185

Brier

2.20x

Top-decile lift

Calibration

Metric panel

ROC AUC: 0.801
PR AUC: 0.712
Brier score: 0.185
Positive rate: 38.5%
Top-decile lift: 2.20x
Temporal test: final four months

Method

LightGBM classifier · 160 trees · learning rate 0.05 · max depth 5 · 24 leaves · balanced class weights · standardized numeric features · one-hot specialty, region and segment.

Regional AUC (sorted)

Gap = 0.149

Review finding. Regional AUC gap = 0.149. This is a review finding, not a hidden metric. Small regional holdouts require confidence intervals, minimum sample thresholds and mitigation before production use.

TIME SERIES + DEEP LEARNING

12-week territory forecasting with a transparent challenger and calibrated uncertainty

Seasonal naive

MAE

7.27

RMSE

9.35

WAPE

20.53%

R²

0.592

LightGBM Poisson

CHAMPION

MAE

5.31

RMSE

7.00

WAPE

15.00%

R²

0.771

LSTM

MAE

5.32

RMSE

7.08

WAPE

15.02%

R²

0.765

National forecast · actual vs models · shaded p10–p90

Show LSTM challenger

Feature importance · top 10

Methodology

Territory-week time split: training through week 79, calibration weeks 80–91, test weeks 92–103.
LightGBM: Poisson objective, 240 estimators, 18 temporal/operational/geographic features.
LSTM: global model, 12-week sequences, 8 normalized inputs, hidden size 28, dense 28→14→1 Softplus head.
Adaptive split-conformal normalized nonconformity quantile q = 0.309.
Nominal 80% interval empirical coverage 86.6%.

Interpretation. LightGBM and LSTM are statistically close. The LSTM challenger improves WAPE by only 0.7% relative, below the pre-defined 2% complexity-aware promotion threshold; LightGBM Poisson therefore remains champion and also has stronger R² and lower operational complexity.

BAYESIAN STATISTICS

Uncertainty-aware response estimates by segment

Posterior mean · 90% credible interval

Access-constrained / maintain37.8% [36.9%, 38.7%]

High-potential omnichannel41.0% [39.9%, 42.1%]

Low-engagement / nurture34.7% [33.6%, 35.8%]

Formula

Beta(1, 1) prior
+ Binomial likelihood
→ Beta(1+y, 1+n-y) posterior

Posterior response probabilities summarize aggregate uncertainty; they do not estimate treatment or engagement uplift.

NLP + CONTROLLED RETRIEVAL

Useful automation without medical-content overreach

5,000

Notes

95.1%

Classifier accuracy

94.9%

Macro-F1

95.0%

Retrieval recall@1

100%

JSON conformance

98.7%

Escalation recall

Confusion matrix

	access_barrier	product_education	logistics	follow_up	safety_question
access_barrier	293	1	4	3	0
product_education	6	308	3	2	2
logistics	8	1	202	4	1
follow_up	6	5	2	270	0
safety_question	2	4	2	5	116

Deterministic RAG flow

1Note classification
2Retrieve approved snippets
3Validate evidence IDs
4Emit structured JSON
5Escalate safety and low-confidence cases

Output contract

{
  "category": "controlled_label",
  "summary": "grounded approved statement",
  "route": "approved_owner",
  "requires_human_review": true,
  "evidence_ids": ["DOC-..."]
}

Guardrails

No diagnosis or treatment recommendation.
No patient-identifying text.
Approved retrieved text only.
Safety language always routed to a human-approved channel.
Prompts, evidence IDs and version retained for audit.

The repository demonstrates a provider-neutral LLM gateway and deterministic fallback without presenting external generated content as validated medical information.

PRESCRIPTIVE ANALYTICS

Capacity-feasible next-best-action planning

Balanced

+7.4% vs matched-random

Budget cap: €34,000
Budget used: €30,000
Field capacity: 110 / 110
Contact cap: 350 / 350
Expected incremental starts: 42.413
Field visits: 12
Hybrid sequences: 98
Remote calls: 142
Approved emails: 0
None: 8

Action distribution

MILP formulation

maximize Σ_i Σ_a utility(i,a) × x(i,a)
subject to one action per HCP
           budget ≤ cap
           field units ≤ capacity
           contacts ≤ cap
           digital actions require consent
           x(i,a) ∈ {0,1}

Warning. Expected uplift coefficients are synthetic assumptions, not causal estimates. Real deployment requires experiments or causal inference before impact claims.

RESPONSIBLE AI

Purpose limitation, evidence and human authority by design

Monitoring

Metric	Value	Threshold	Status
Contacts PSI	0.00001692	0.2	PASS
Forecast WAPE	0.1500	0.18	PASS
Propensity Brier score	0.1853	0.2	PASS
Interval coverage	0.8663	0.75	PASS

Risk register

Purpose creep

Low with enforced intended/prohibited uses.

PASS

Personal-data leakage

Low in synthetic project; enterprise DLP required.

PASS

Regional performance disparity

AUC gap 0.149 across regions.

REVIEW

LLM hallucination

Prototype/review — approved retrieval only.

REVIEW

Automation bias

Ongoing oversight, no auto-send.

HUMAN REVIEW REQUIRED

Invalid causal interpretation

Review before impact claims.

REVIEW

Human oversight

Propensity

Commercial owner may override or exclude.

Human oversight

Forecast

Planner compares champion, challenger, interval and local events.

Human oversight

NLP / RAG

Safety and low-confidence items escalate; no automatic sending.

Human oversight

Optimization

Plan is reviewed for feasibility, consent and local context.

Governance is designed to be compatible with GDPR/CNIL risk-management expectations and the progressive EU AI Act lifecycle. This synthetic case study does not claim a formal legal classification or compliance opinion.

PRODUCTION ENGINEERING

A deployable model package, not a notebook-only demonstration

End-to-end fixed-seed pipeline

src/pipeline/build_all.py

PySpark medallion transformations

src/data/spark_medallion_pipeline.py

Databricks notebook pattern

databricks/01_medallion_pipeline.py

FastAPI typed scoring

src/api/main.py

Docker image

Dockerfile

GitHub Actions CI

.github/workflows/ci.yml

Model registry & model cards

artifacts/registry, artifacts/model_cards

Terraform cloud scaffold

infra/terraform

Contract tests (6 passing)

tests/test_contracts.py

Power BI assets (DAX + dark theme)

dashboards/powerbi

Tableau workbook

dashboards/tableau/control_tower.twb

Excel control tower (12+ sheets)

workbooks/control_tower.xlsx

API contract

GET  /health      → status + model_version
POST /propensity  → typed features → probability + decision_support_only=true

Build

Reproducible fixed-seed pipeline
Semantic model version
Data + feature contracts

Validate

Temporal holdouts
Calibration + conformal coverage
Fairness by region

Operate

Monitoring dashboards
PSI drift + coverage checks
Rollback via registry pin

REPRODUCIBILITY

Every headline metric has a downloadable evidence trail

Excel control workbook

12+ sheets, formulas and reconciliations.

Download .xlsx

Technical report

23-page methodology, results, limitations and sources.

Read PDF

Complete model package

Code, data, models, tests, dashboards and infrastructure.

Download .zip

GitHub repository

Source and version history.

Open GitHub

Workbook architecture

README→

Parameters→

Dashboard→

Forecast→

Scores→

Segments→

NBA plan→

Bayesian→

Data quality→

RAI register→

Data dictionary→

Sources→

Scenarios

Green workbook cells are formulas, yellow cells are Python/model outputs and blue cells are editable assumptions.