Industry leader — anomaly detection for serial production (in progress)

Context

The client is a Polish serial-production leader with a plant in central Poland. The production area has 50 people across four process lines. Each line runs its own legacy SCADA, built by three different vendors over the last 10 years — a typical technology tail for a mid-size Polish manufacturer, where tool decisions followed capex cycles, not architecture.

The plant runs three shifts, non-stop. The board contracted Growth Platinum AiP. The work is two phases: the AiP Audit (3 months, complete) and QDeployment Performance (14 months of build, in progress). Last-line go-live is planned for 2026 Q3.

The board came with a business question, not a technology question: why plant TDC sits at 71% after successive hardware upgrades, and where the throughput points are actually lost. That answer needed an audit, not a deployment, so the engagement opened on Growth Platinum, not QDeployment.

Challenge

The AiP Audit found four linked problems whose cost the board had not aggregated.

First — anomalies caught after the fact, with a 4–8 hour delay. A line operator noticed a precision drop mid-shift, stopped the batch, and the batch was already a write-off. Unit loss: 8–40 thousand zloty depending on batch size.

Second — no central root-cause diagnosis. Two shifts could reach different conclusions on the same defect. Corrective calls sat with the shift lead, on feel and shift experience, not data. Each miss cost another 2–3 batches before the hypothesis was tested.

Third — TDC (Throughput-to-Defect Coefficient) reported as a plant-wide aggregate at 71%. The board did not know which process stage leaked the points. No TDC split per line meant capex for upgrades could not be allocated rationally.

Fourth — SCADA → ERP integrations were a nightly batch with a 12-hour lag, and the quarterly board report was written by hand by a process engineer in the three days before the presentation.

The first three weeks of the AiP Audit surfaced the number that closed the Performance conversation: 63% of the process engineer’s time went to manually correlating SCADA events with batch parameters — work an algorithm can do in seconds. That one metric convinced the board to contract Growth Platinum with an immediate path to QDeployment Performance for all four priority processes.

Approach

01 — Growth Platinum AiP Audit (3 months, complete). Process mining on SCADA, ERP and MES logs from 12 months of history. Activity-based costing on four priority processes. Cross-check by two independent auditors (Platinum standard, countersignature required). Audit result: four processes to automate, designed NPV per process 270–1080 thousand zloty, designed payback 5.8–17.2 months, board countersignature on Performance for all four processes in one contract.

02 — Multi-cloud, multi-model architecture (1 month, designed). The first architectural decision that saves the client millions — we do not unify SCADA. Three legacy vendors stay as data sources; we add an event-stream layer above them. AWS Bedrock plus Anthropic Claude for anomaly detection on time series — trained on 18 months of history, fine-tuned on defect signatures for this portfolio. Google Vertex AI plus Gemini 3.1 Pro for root-cause hypotheses with process-condition context. Legacy SCADA integration through an OPC UA bridge plus a custom MQTT broker. Compliance: AWS DPF and EU-US SCCs, sensitive data stays in the client VPC, model fine-tune offline.

03 — QDeployment Performance build (10 months, in progress). Phase one (5 months) — pilot on one line, event-stream plus anomaly detection plus operator dashboard. Currently in UAT on the pilot line. Phase two (5 months) — rollout to the remaining three lines plus root-cause engine plus automated quarterly board reporting. Gate A after the Audit (architecture countersignature — passed), Gate B after the pilot (TDC acceptance on the pilot line — in progress), Gate C after UAT on line one (green light for rollout — planned Q3 2026).

04 — 3 weeks of hypercare and QCare Performance handoff (planned Q3 2026). TRL 9 SLA — the system in full production, not a pilot. After last-line go-live, three weeks of intensive hypercare with a dedicated SRE and daily standups. Weeks 20–24 post-go-live: handoff to the QCare Performance subscription.

Stack

Multi-cloud, multi-model, production-shaped — designed in phase 02, now in build:

• AWS Bedrock (eu-central-1) plus Anthropic Claude — real-time anomaly detection on the SCADA event stream
• Google Vertex AI plus Gemini 3.1 Pro — root-cause hypothesis generation with 4-vendor context
• OPC UA bridge plus MQTT broker (custom) — three SCADA vendors
• Snowflake (eu-central-1) — data warehouse, 18 months retention
• Looker Studio plus custom dashboards — operator (per line) plus board (quarterly ROI)
• NVIDIA Inception Program — access to preview models tuned for time-series
• Compliance: AWS DPF plus EU-US SCCs, sensitive data stays in the client VPC, anomaly model fine-tune offline

Every stack choice has an ADR (Architecture Decision Record) with rejected alternatives and cost of change — a standard QDeployment deliverable, not an add-on.

Designed effects

Numbers from the audit — values calculated in phase 01 (Growth Platinum AiP Audit) as the basis of the QDeployment Performance contract. To be verified in a 90-day cycle after go-live, quarterly report under QCare.

Metric	Pre-deployment	Designed target	Delta
TDC (Throughput-to-Defect Coefficient)	71%	89%	+18 p.p.
Yearly benefit hidden in repeatable processes	—	712 000 zł	—
One-off investment (Audit + QDeployment)	—	575 000 zł	—
Designed payback	—	≈14 mo.	—
Hours freed in the team (yearly)	—	2 592 h	≈ 1.3 FTE
Anomalies caught in real time	0%	87%	+87 p.p.
Time to detect an anomaly	4–8 h	<8 min	-98%
Quarterly reporting time	3 days	4 h	-94%
Rejected batches (yearly)	142	38	-73%

The 712 000 zł yearly benefit hidden in repeatable processes is the sum of the four processes under the QDeployment contract. The largest line is designed elimination of rejected batches — real-time anomaly detection cuts about 104 batches a year at an average 13 thousand zloty per batch, which is 1.35 million zł gross, or about 685 000 zł net after model and platform cost. The remaining 27 000 zł is designed saving on the quarterly report — three process-engineer days times four quarters times fully loaded rate.

What next

After go-live on all four lines and hypercare (planned Q3 2026) the project moves to the QCare Performance subscription — operational maintenance with SLA, a quarterly ROI report and a 12-month renegotiation cycle. The table values will be verified in the first reporting cycle 90 days after go-live.