01 The Challenge
Koch Industries operates across a wide span of business units — energy, manufacturing,
materials, consumer products, and more. Production planning across that portfolio depends
on getting accurate, timely forecasts of production capacity: how much
of what can be produced, where, and when, given the current state of the underlying
feature data.
Two compounding problems made this hard. First, the feature-data pipeline was a
throughput bottleneck — the inputs to the ML models were not arriving fast
enough or in the right shape for forecasts to stay current with reality. Second, even
when the data was available, the forecasts needed to be served as operational
APIs that downstream business systems and decision-makers across multiple business
units could consume directly — not as one-off notebooks or reports.
02 The Approach
Leopard Data was originally engaged to parallelize the feature-data processing
— to take an inherently slow, sequential preparation pipeline and re-architect it for
horizontal scale on a distributed compute cluster. That work succeeded and the engagement
expanded.
With the throughput problem solved, we moved into the broader platform: REST services
on top of an ensemble of AI / ML forecasting models, served from a Kubernetes
cluster on AWS EKS with KEDA event-driven autoscaling so the inference layer could scale
with demand from each Koch business unit. The model portfolio included statistical baselines
(ARIMA, Holt-Winters), modern probabilistic time-series methods (Prophet, DeepAR), and
deep recurrent models (LSTM) — each chosen and tuned for the type of production data
it was forecasting.
The result was an internal platform that any Koch business unit could call: same API
surface, same operational guarantees, same forecasting discipline, regardless of which
business’ production capacity it was being asked about.
03 The Outcome
- Feature-data processing pipeline parallelized; the throughput bottleneck eliminated.
- AI / ML production-capacity forecasting platform deployed and serving multiple Koch business units via REST.
- Model ensemble in production: ARIMA, Holt-Winters, Prophet, LSTM, and DeepAR — matched per data shape.
- Operational on AWS EKS with KEDA autoscaling against per-unit forecast demand.
04 Tech Stack
Python
Ray / Anyscale
AWS EKS
Kubernetes
KEDA
PostgreSQL
REST APIs
ARIMA
Holt-Winters
Prophet
LSTM
DeepAR
