CTO · S-01
Why two separate ML models instead of one unified model?
"You have a RUL regressor for batch and an Isolation Forest for streaming. Why not one model that does both? The operational complexity of two separate Vertex AI endpoints, two monitoring jobs, and two HITL paths seems high."
Architectural response
The two models have fundamentally different problem formulations that cannot be merged without compromising both. The RUL regressor is a supervised regression model — it requires historical failure labels and produces a continuous prediction (days to failure) with a confidence interval. It is designed for daily batch processing where the full feature history of each unit is available. The Isolation Forest is an unsupervised anomaly detector — it requires no failure labels and produces a binary signal (normal vs abnormal) based on real-time sensor deviation from a learned baseline. It is designed for streaming, event-driven processing where a single sensor reading triggers an immediate response. Merging them would require a semi-supervised model that is harder to train, harder to explain, and harder to monitor. Separation of concerns is the right architectural decision: the RUL model answers "when will this unit fail?" on a planning horizon; the anomaly model answers "is something wrong with this unit right now?" on an operational horizon.
Evidence: ADR-AQ01 (different compute substrates) · Page 06 Model 02 (RUL) and Model 03 (Anomaly) — separate model cards, separate evaluation metrics, separate confidence thresholds
CCO · S-02
Does the RUL prediction fall under EU AI Act Annex III?
"The EU AI Act Annex III covers AI systems used in critical infrastructure. Does a predictive maintenance model for MRI scanners in hospitals qualify? And if so, what are the specific documentation obligations?"
Architectural response
Yes. EU AI Act Annex III Category 2 covers AI systems used in the management and operation of critical infrastructure — including healthcare facilities. An MRI scanner failure in a hospital is a patient safety event. A predictive maintenance model that determines when maintenance interventions occur is directly influencing the operational status of medical critical infrastructure. The documentation obligations are: Article 11 (Technical documentation — satisfied by the Asset IQ Model Card on Page 06), Article 13 (Transparency — satisfied by SHAP sensor attribution per prediction), Article 14 (Human oversight — satisfied by HITL-06 for low-confidence predictions and HITL-07 for fleet anomalies), and Article 15 (Accuracy and robustness — satisfied by the drift detection monitoring and the quarterly full-rebuild validation). The ISO 13485 DHR requirement is separate from EU AI Act — it applies to the Device History Record for the physical MRI unit, not to the AI model itself.
Evidence: Page 06 Asset IQ Model Card · HITL-06/07 specification · EU AI Act Annex III Cat. 2 · ISO 13485 DHR integration in state machine
Field Service Manager
What if the model predicts a failure that doesn't happen?
"If Asset IQ sends me to a site for a planned intervention and the unit is fine, I've wasted an engineer's day. At €800 per field visit, false positives are expensive. What's the false positive rate and who is accountable when it's wrong?"
Architectural response
The false positive question is answered by design: confidence below 0.82 routes to you for a decision — it does not auto-dispatch an engineer. When the model is uncertain, you decide. At 0.82 confidence and above, the precision at the 14-day horizon is 0.87 — meaning 13% of auto-dispatched interventions are for units that would not have failed within 14 days. At approximately €800 per visit versus approximately €42,000 for an emergency dispatch plus hospital disruption costs, the break-even false positive rate is well above 13%. When the model is wrong, the override you enter in HITL-06 becomes a training example that improves the model's precision on that failure mode. Accountability is shared by design: below-threshold predictions require your professional judgment before any action is taken.
Evidence: Page 06 Asset IQ RUL Model Card (Precision@14d: 0.87, confidence threshold 0.82) · HITL-06 three-decision interface · override as training signal (concept drift section Page 06)
Enterprise Architect · S-08
How do you handle 6 regional systems with different sensor schemas?
"The 6 regional asset management systems were built at different times by different vendors. Their sensor data schemas are almost certainly inconsistent. How does Asset IQ produce a unified feature vector from six incompatible data sources?"
Architectural response
ADR-AQ02 is the direct answer to this question. The unified Pub/Sub canonical schema is the architectural decision that resolves schema inconsistency at the ingestion boundary. Each regional system has a lightweight ingestion adapter that translates its native schema to the canonical asset event schema before publishing to Pub/Sub. The Data Governance module (M-08, H1) validates every Pub/Sub message against the canonical schema — malformed records are quarantined before they reach the Feature Store. The canonical schema defines 24 mandatory sensor fields — regional systems that do not produce all 24 fields publish null values for missing sensors, which are imputed at feature engineering time using fleet median values. The RUL model was trained on this imputed feature matrix — it handles missing sensors gracefully without requiring data quality perfection from every regional system on day one.
Evidence: ADR-AQ02 (unified schema) · M-08 Data Governance (schema validation gate) · Page 06 Asset IQ Feature Store (24 features, null imputation strategy)
CISO · S-09
Sensor telemetry data — what data classification applies?
"The asset telemetry includes scanner operating data from hospitals. Does that constitute health data under GDPR? Or is it purely machine data with no personal data implications?"
Architectural response
The sensor telemetry processed by Asset IQ is machine operational data — coil temperatures, helium levels, RF power deviation, scan utilisation rates. It does not include DICOM patient images, patient identifiers, scan schedules, or any data that can be linked to an individual patient. The data classification is Internal (not Confidential or Restricted) — it is commercially sensitive (ClaraVis's operational performance data) but not personal data under GDPR. The Pub/Sub canonical schema explicitly excludes any patient-linked fields — this is validated by the Data Governance schema at ingestion. The DLP API is not configured for asset telemetry because there is no PII to detect. The EU boundary (europe-west3 VPC-SC) applies to all asset telemetry regardless of data classification — both for data sovereignty and because the Org Policy region constraint applies uniformly across the entire GCP project.
Evidence: Page 07 data classification labels (asset telemetry: internal) · Pub/Sub canonical schema (no patient fields) · Page 07 DLP configuration (contracts only, not telemetry)
VP Field Service · S-06
Can we use RUL predictions to renegotiate warranty reserves?
"We currently hold €40M in warranty reserves because we assume worst-case failure scenarios. If Asset IQ can predict failures with 87% precision at 14 days, can the CFO use those predictions to reduce the reserve? And what's the audit trail if the predictions are wrong?"
Architectural response
This is a financial accounting decision that requires input from the Finance Controller and auditors — Asset IQ provides the data, not the accounting judgment. What Asset IQ can provide: a fleet-level RUL distribution showing the probability of failure within each time window across all 12,000 units, with confidence intervals. This distribution, combined with historical actual failure data and the precision metrics from the Model Card, gives the Finance team a defensible actuarial basis for reserve calculation that is more accurate than worst-case assumptions. The audit trail is already designed for this: every RUL prediction is written to the BigQuery Device entity with confidence interval, SHAP attribution, and the model version that produced it. The HITL decision (human override or approval) is linked to the prediction record. If an auditor asks "what was the fleet failure probability distribution on 31 March 2026?", the answer is a BigQuery query returning 12,000 rows of predictions with confidence intervals — an evidence package that justifies any reserve calculation based on it.
Evidence: BigQuery Device entity (RUL score · confidence interval · model version) · HITL override records (linked to prediction) · Page 06 Asset IQ Model Card (precision metrics for audit defence)