JOHN HOLSTEIN — ASCEND PERFORMANCE MATERIALS

Someone with chemicals experience knows the molecules. I know the supply chain physics that governs those molecules — batch manufacturing, hazmat transport, JIT constraints — regardless of what they are. The ADN global bottleneck (three large-scale producers worldwide) is structurally identical to the Mo-99 reactor concentration risk I managed at Cardinal Health for years. ML architecture and supply chain judgment take years to build; the chemistry I learn in 90 days, and it is the least complex part of this role.

Post-emergence is the highest-leverage moment to architect a data science function from first principles. The balance sheet restructuring removed a structural overhang — $1.3B in debt eliminated, $600M in new capital. The operational assets were never impaired. What the restructuring created is a clean-sheet moment: a new CEO who named specific AI use cases publicly on day one, and cost discipline that forces ROI-first model development, not research theater. I have done this build at Cardinal Health, Invistics, and J&J. I know what the first 90 days must produce.

Payback speed times certainty of outcome — not NPV over five years. The company needs wins visible within one quarter. AR first: fastest cash impact on ABL availability, and a working propensity model in 8 weeks (4 weeks SAP FI/CO data architecture, 4 weeks model training). PO automation second: Conexiom-to-Celonis extension build, lower execution risk. Supply chain sequencing third: highest complexity across SAP and DeltaV, earns its investment once the first two wins are banked and credibility is established.

Payment propensity scoring: which customers will pay late, by how many days, and what is the expected cash impact. Feature set: historical days-past-due by account, invoice size, payment behavior by product line, customer end-market seasonal patterns. Segment by customer type — automotive Tier 1s have different behavior profiles than electronics OEMs or apparel mills. At a company that just negotiated a $350M ABL facility, the difference between an accurate 30-day cash forecast and an inaccurate one directly affects whether you draw on the revolver unnecessarily. That is a CFO-visible outcome.

At J&J MedTech, I aligned 23 independent business units — each with its own IT governance, finance leadership, and SAP configuration — around a common data architecture with zero direct authority over any of them. I started with a read-only schema mapping exercise, not a data migration. No one's system was being touched. The data dictionary I published became the reference the units adopted themselves, and the remaining 20 followed without requiring escalation once three were voluntarily contributing. That S3 data lake became the foundation for all subsequent AI work.

Benzene from crude oil refinery operations converts to cyclohexane, which oxidizes to KA oil (cyclohexanol + cyclohexanone), then oxidizes to adipic acid — feedstock cost tracks crude and benzene spot prices directly, so your adipic acid margin is partly an oil price derivative. Separately, acrylonitrile (also crude-linked via propylene) goes through Ascend's electrochemical hydrodimerization process to ADN, then hydrogenates to HMD. The ADN bottleneck — three large-scale global producers, two process routes, no easy substitution — mirrors the Mo-99 reactor concentration risk I managed at Cardinal Health. That supply chain architecture I know. Internal unit economics at each node: I build in the first 30 days.

The target variable lives in SAP QM, in the batch disposition record. Leading indicators live in DeltaV historian — temperature profiles, pressure, residence time, catalyst concentration over the reaction window. The bridge is an event-aligned join keyed on SAP batch number to DeltaV time window; that key is built from production order completion timestamps in SAP PP. I did this exact schema archaeology across 23 J&J configurations — I know where those joins live. Feature engineering runs on the historian side. Azure Data Factory lands cleaned extracts into the analytics environment. Output: a batch release risk score quality teams act on before lab results, not after. That is the payback.

Multivariate time series layering OEM production schedules (published quarterly, updated monthly) against historical Ascend HMD order patterns, with benzene and cyclohexane spot price as feedstock cost covariates. Critical architectural decision: the ADN supply-side constraint requires a disruption indicator detecting supply shock before it propagates to HMD availability — at Invistics I built the equivalent logic for hospital supply chains where the constraint was distributor allocation limits. The output is a probability distribution over demand scenarios, not a point forecast — the supply planning team needs to act on the distribution, with dynamic safety stock calibrated to the 3-4 week lead time distribution.

Three layers. Data foundation: unified supply chain event store ingesting from SAP PP production orders, SAP MM goods movements, DeltaV historian for real-time plant status, and AVEVA BI Gateway for connected worker data. Model layer: network flow optimization for deterministic daily planning (ADN from Decatur and adipic acid from Pensacola as the co-binding constraints), with Monte Carlo simulation for disruption scenarios calibrated to the Wilson Lock closure, Pensacola fire, and Texas freeze as the three calibration events I already have. Action layer: Celonis Action Flows surfacing outputs to supply planners. Critical design decision: 4-hour intraday ETL refresh cycle — monthly batch makes sequencing optimization operationally irrelevant.